Benchmarks
More
OverviewBenchmaxxingCompareChartsCost vs CapabilityProviders per ModelProvider explorerGatewaysEU & SovereignJev-class modelsImage JevBenchAudioJevBenchSubmit a modelAbout
JevBench · method v1.8 (10 s rule, GPU-time cost basis) · updated 2026-11-03

JevBench — decision model benchmark & leaderboard

A benchmark for AI decision models: state and rubric in, typed answer out. We compare accuracy, calibration, latency and cost, independently of TypeSafe AI.

235 ranked open-weights systems · Jev 1.13.0 and Jev 1.14.0 shown as references · 1,500 decisions each · 110 newer ranked rows tagged with their draw · 13 older rows dated separately · How it works · Release history · Methodology · Data & JSON/CSV · aggregate JSON · Share · Other decision benchmarks · Explore ImageJevBench v0.3.0

Speed and cost are compared within a group: open weights on the same GPU, APIs as sold by each provider.

JevBench board · headline

JevBench Capability Score (open weights)

Capability ranking of Jev-class systems

Capability Score averages Intelligence and Calibration. baselabs mammoth-1 leads the Jev-class systems with 85.2.

Jev-class means at most 2× the cost and median latency of Jev (v1.5 reference). How we choose ↘

Adjust cost / latency caps · 2× official
Official
#SystemICCap.$/1k
  1. 1baselabs mammoth-1v1.7.0 drawv1.7 · 10 s rule76.257.285.2$0.027*
  2. 2KnowLine 31B-Gen1v1.7.0 drawv1.7 · 10 s rule77.446.585.2$0.061*
  3. 3Fastino GLiDE 1.2v1.7.0 drawv1.7 · 10 s rule76.458.184.9$0.025*
  4. 4H2O-Lightning-31B v1.0.7v1.6.43 drawv1.7 · 10 s rule79.356.084.3$0.029*
  5. 5Perplexity Decider v1.1 27Bv1.7 · 10 s rule75.657.382.7$0.026*
  6. 6vLLM-SR d3-turbov1.6.43 drawv1.7 · 10 s rule74.857.382.6$0.027*
  7. 7StartLux-Decision v1.1 27B Instantv1.6.43 drawv1.7 · 10 s rule73.257.982.6$0.025*
  8. 8vLLM-SR d3v1.6.43 drawv1.7 · 10 s rule72.957.382.1$0.027*
  9. 9basal-omni-1.0-ultrav1.6.43 drawv1.7 · 10 s rule74.155.982.1$0.029*
  10. 10DOPEST 1.0 31Bv1.7.4 drawv1.7 · 10 s rule73.952.682.0$0.038*
  11. –Jev 1.13.0API63.654.777.1$0.032Not ranked, only shown as a reference to compare with
  12. –Jev 1.14.0API59.954.875.0$0.032*Not ranked, only shown as a reference to compare with
Show all 219 Jev-class systems (207 more)
  1. 11Jeff-1.1-Mediumv1.7.8 drawv1.7 · 10 s rule74.566.281.9$0.013*
  2. 12Arcbright Elf 1 31Bv1.7.0 drawv1.7 · 10 s rule71.955.181.8$0.031*
  3. 13Quyet-1.0-Largev1.7 · 10 s rule73.455.981.7$0.030*
  4. 14Jeffan-31b-1710v1.7.0 drawv1.7 · 10 s rule71.651.481.7$0.042*
  5. 15Jeff-1.0-Largev1.6.4 drawv1.7 · 10 s rule71.954.581.5$0.033*
  6. 16Nimble-v4-maxv1.7.0 drawv1.7 · 10 s rule71.057.581.2$0.026*
  7. 17Sieve-31B-v2v1.7.5 drawv1.7 · 10 s rule68.756.281.1$0.029*
  8. 18Turion Jev 31Bv1.7.1 drawv1.7 · 10 s rule70.946.680.9$0.060*
  9. 19pplx-decider-v1.2v1.6.43 drawv1.7 · 10 s rule71.757.480.8$0.026*
  10. 20torchcast-decision-27bv1.7 · 10 s rule70.457.780.2$0.026*
  11. 21Lychee-27Bv1.7.20 drawv1.7 · 10 s rule74.357.379.8$0.026*
  12. 22metask-jev-rain-31B-trained service v3v1.6.43 drawv1.7 · 10 s rule72.655.479.6$0.031*
  13. 23decisio v0.8.0 on gemma-4-31B-itv1.7 · 10 s rule70.455.079.6$0.032*
  14. 24Vera Largev1.7.0 drawv1.7 · 10 s rule67.855.979.6$0.029*
  15. 25Kev 27Bv1.7 · 10 s rule64.256.278.9$0.029*
  16. 26metask-jev-rain-31B-trainedv1.6.43 drawv1.7 · 10 s rule70.756.178.9$0.029*
  17. 27StartLux-Decision v1.1 27B AutoThinkv1.6.43 drawv1.7 · 10 s rule66.756.678.8$0.028*
  18. 28Bobcat 1.3v1.7.0 drawv1.7 · 10 s rule68.155.277.9$0.031*
  19. 29deck-31Bv1.7 · 10 s rule73.055.077.6$0.032*
  20. 30Torchcast Decision 4Bv1.6.43 drawv1.7 · 10 s rule67.179.477.5$0.0049*
  21. 31KnowLine 26B-A4B-Gen2.5v1.7.18 drawv1.7 · 10 s rule65.262.377.3$0.018*
  22. 32Eikos-27Bv1.7 · 10 s rule64.456.177.2$0.029*
  23. 33metask-jev-rain-12B-trainedv1.6.43 drawv1.7 · 10 s rule66.565.276.9$0.014*
  24. 34Wald 26B-A4Bv1.7.0 drawv1.7 · 10 s rule65.571.076.8$0.0093*
  25. 35KnowLine 35B-A3B-Gen1.5v1.7.0 drawv1.7 · 10 s rule60.363.676.7$0.016*
  26. 36René-1 31B FP8v1.7 · 10 s rule61.751.376.0$0.042*
  27. 37Decisio 12B tuned commitv1.7.6 drawv1.7 · 10 s rule62.867.075.7$0.013*
  28. 38Clefv1.7 · 10 s rule59.155.575.3$0.030*
  29. 39H2O-Lightning-4B v1.1v1.6.43 drawv1.7 · 10 s rule60.078.675.0$0.0052*
  30. 40JEV-27Bv1.7 · 10 s rule56.858.673.9$0.024*
  31. 41Nagi-ENORMOUS 27Bv1.6.43 drawv1.7 · 10 s rule62.457.473.8$0.026*
  32. 42Surogate Rune 26B-A4B v3v1.7 · 10 s rule56.171.073.5$0.0092*
  33. 43RSI-Jev v6.1-VL 27Bv1.6.9 drawv1.7 · 10 s rule61.658.773.5$0.024*
  34. 44Bobcat Flash 1.2v1.7 · 10 s rule65.671.173.0$0.0092*
  35. 45Torchcast Decision 4Bv1.6.43 drawv1.7 · 10 s rule60.379.472.6$0.0049*
  36. 46swanOnev1.7 · 10 s rule56.271.472.4$0.0090*
  37. 47FrontiersMind Lumma-fev-27Bv1.7.0 drawv1.7 · 10 s rule55.554.872.3$0.032*
  38. 48torchcast-decision-12bv1.7 · 10 s rule60.556.471.7$0.028*
  39. 49AutoJev-27Bv1.7 · 10 s rule55.656.771.6$0.028*
  40. 50decider-12b v2v1.7 · 10 s rule61.968.271.3$0.012*
  41. 51Decision 2.0 Vega 27Bv1.7 · 10 s rule60.457.071.3$0.027*
  42. 52Jev-Omniv1.7 · 10 s rule55.566.571.3$0.013*
  43. 53Winnow-12B Q8v1.7 · 10 s rule59.567.071.2$0.013*
  44. 54Xor 26B-A4Bv1.7 · 10 s rule57.572.171.2$0.0085*
  45. 55decider-12b v1, stock Gemma-4-12B-itv1.7 · 10 s rule59.668.271.0$0.012*
  46. 56Cygnetv1.7 · 10 s rule54.866.870.9$0.013*
  47. 57decisio v0.8.0 on gemma-4-12B-itv1.7 · 10 s rule52.367.770.7$0.012*
  48. 58JPT-35B-A3Bv1.7 · 10 s rule51.472.570.7$0.0083*
  49. 59decider chat on Gemma-4-31B-itv1.7 · 10 s rule58.556.170.4$0.029*
  50. 60baselabs elephant-1v1.7.0 drawv1.7 · 10 s rule55.868.170.2$0.012*
  51. 61Jebadiah 27Bv1.7 · 10 s rule57.557.570.2$0.026*
  52. 62GEV-26B-Decidev1.7 · 10 s rule49.872.670.1$0.0082*
  53. 63Chon-1.0-12B-Fastv1.7.1 drawv1.7 · 10 s rule60.369.170.1$0.011*
  54. 64decisio v0.9.0 on gemma-4-12B-itv1.7 · 10 s rule52.367.770.0$0.012*
  55. 65Blink v0.3 26B-A4Bv1.6.9 drawv1.7 · 10 s rule49.271.169.9$0.0092*
  56. 66NInfer Qwen3.8-Flash-Next mixedv1.7 · 10 s rule49.071.869.7$0.0087*
  57. 67Respan Span-02v1.6.43 drawv1.7 · 10 s rule60.953.569.3$0.035*
  58. 68Philly Decider v2v1.7.10 drawv1.7 · 10 s rule56.179.169.3$0.0050*
  59. 69Wald 4B v2.1v1.6.9 drawv1.7 · 10 s rule52.060.869.3$0.020*
  60. 70JevOnev1.7 · 10 s rule46.072.868.9$0.0081*
  61. 71vLLM-SR d3-miniv1.7.0 drawv1.7 · 10 s rule50.878.868.8$0.0051*
  62. 72Torchcast Decision 4Bv1.6.43 drawv1.7 · 10 s rule54.279.468.7$0.0049*
  63. 73iota-decide-4b-v0.3v1.7.10 drawv1.7 · 10 s rule52.578.968.6$0.0051*
  64. 74Xevn-4Bv1.7.0 drawv1.7 · 10 s rule52.180.568.6$0.0045*
  65. 75Chon-1.0-E4Bv1.7.3 drawv1.7 · 10 s rule55.583.568.5$0.0035*
  66. 76Hopper 12B trainedv1.7 · 10 s rule51.766.668.1$0.013*
  67. 77SPX-CD-Omniv1.7 · 10 s rule48.868.768.1$0.011*
  68. 78NInfer Qwen3.8-27B NVFP4v1.7 · 10 s rule51.856.568.0$0.028*
  69. 79Aplomb 1v1.7 · 10 s rule45.272.467.1$0.0083*
  70. 80Bespoke Nimble 9B v3v1.7 · 10 s rule56.565.866.8$0.014*
  71. 81Amazon Strands Decider 26B-A4Bv1.7.1 drawv1.7 · 10 s rule46.771.966.7$0.0086*
  72. 82Tern-35B-A3Bv1.7.0 drawv1.7 · 10 s rule44.171.366.4$0.0090*
  73. 83KnowLine 4B-Gen5.5v1.7.18 drawv1.7 · 10 s rule46.469.866.2$0.010*
  74. 84decider-35b-a3bv1.7 · 10 s rule48.773.266.2$0.0078*
  75. 85Metask rain 4Bv1.6.4 drawv1.7 · 10 s rule51.169.566.0$0.010*
  76. 86shisa-de-1v1.6.43 drawv1.7 · 10 s rule49.971.165.5$0.0092*
  77. 87Plumb-4Bv1.7 · 10 s rule42.877.665.3$0.0056*
  78. 88Arcbright Elf 1 12Bv1.6.43 drawv1.7 · 10 s rule43.867.865.0$0.012*
  79. 89decider-4b v2v1.7 · 10 s rule40.179.164.9$0.0050*
  80. 90Seb-9B-v2v1.7.0 drawv1.7 · 10 s rule46.168.664.8$0.011*
  81. 91JevK5 v0.3v1.7 · 10 s rule38.577.664.2$0.0056*
  82. 92Jeffan-4B-2010v1.7.0 drawv1.7 · 10 s rule40.178.764.0$0.0051*
  83. 93Clef-Flashv1.7 · 10 s rule42.266.563.8$0.013*
  84. 94JPT-4Bv1.6.43 drawv1.7 · 10 s rule40.878.963.8$0.0051*
  85. 95deck-4B v1.0v1.7 · 10 s rule38.678.063.6$0.0054*
  86. 96Jebadiah 9B v2.1v1.6.43 drawv1.7 · 10 s rule37.268.563.4$0.011*
  87. 97GLiClass v4 9Bv1.7.9 drawv1.7 · 10 s rule39.670.963.2$0.0093*
  88. 98Seb-9Bv1.6.43 drawv1.7 · 10 s rule38.368.663.1$0.011*
  89. 99janus 4Bv1.7 · 10 s rule37.679.062.9$0.0050*
  90. 100JevK5 v0.2.0v1.7 · 10 s rule36.377.662.8$0.0056*
  91. 101Messier One v0.4v1.6.43 drawv1.7 · 10 s rule52.075.762.4$0.0065*
  92. 102Imajev-9Bv1.6.43 drawv1.7 · 10 s rule36.750.562.4$0.045*
  93. 103Imajev-4Bv1.7 · 10 s rule34.677.861.9$0.0055*
  94. 104jqvv1.7 · 10 s rule33.954.761.9$0.032*
  95. 105metask-jev-4bv1.7 · 10 s rule39.276.061.8$0.0063*
  96. 106Kahn1 4Bv1.6.43 drawv1.7 · 10 s rule39.764.761.7$0.015*
  97. 107pplx-decider-flash-v1.2v1.6.43 drawv1.7 · 10 s rule45.579.061.7$0.0050*
  98. 108levv1.7 · 10 s rule41.176.361.7$0.0061*
  99. 109Decision 2.0 Lux 9Bv1.6.43 drawv1.7 · 10 s rule35.868.161.6$0.012*
  100. 110JEB-35B-A3Bv1.6.43 drawv1.7 · 10 s rule33.272.361.6$0.0084*
  101. 111Decisio v0.9.0 on Qwen3.6-35B-A3Bv1.7.10 drawv1.7 · 10 s rule34.061.761.5$0.019*
  102. 112blink-4bv1.6.43 drawv1.7 · 10 s rule35.078.161.4$0.0054*
  103. 113Hopperv1.7 · 10 s rule33.676.861.2$0.0059*
  104. 114Malkuth-4Bv1.7 · 10 s rule33.877.561.1$0.0056*
  105. 115Qwen3.5-9B Jev-like data-mix v2v1.7 · 10 s rule41.965.460.9$0.014*
  106. 116Decision 4B v1.2v1.7 · 10 s rule32.477.660.4$0.0056*
  107. 117jiwo 4bv1.6.9 drawv1.7 · 10 s rule34.878.660.2$0.0052*
  108. 118Verdin-4Bv1.7.0 drawv1.7 · 10 s rule35.476.759.6$0.0060*
  109. 119Wald 4Bv1.6.4 drawv1.7 · 10 s rule41.070.559.6$0.0096*
  110. 120Quyet-1.0-Mediumv1.7 · 10 s rule41.778.059.2$0.0054*
  111. 121Decision-4Bv1.7 · 10 s rule37.980.459.1$0.0045*
  112. 122Clef-omniv1.6.9 drawv1.7 · 10 s rule34.556.058.3$0.029*
  113. 123Manchego v2.1v1.7 · 10 s rule32.578.958.3$0.0051*
  114. 124Messier One v0.2v1.7 · 10 s rule42.479.458.1$0.0048
  115. 125Diffusion Jevv1.7 · 10 s rule54.971.258.1$0.0091*
  116. 126jev-localv1.7 · 10 s rule41.978.557.5$0.0052*
  117. 127Decision 4B v1.1v1.7 · 10 s rule29.177.657.2$0.0056*
  118. 128JEV Qwen3.5-9B Base NVFP4v1.7 · 10 s rule29.347.657.1$0.056*
  119. 129iota-decide-4bv1.6.43 drawv1.7 · 10 s rule29.477.557.1$0.0056*
  120. 130Uniphore-4B-Deciderv1.6.43 drawv1.7 · 10 s rule34.279.157.1$0.0050*
  121. 131Jebadiah 9B v2v1.6.43 drawv1.7 · 10 s rule30.668.556.9$0.011*
  122. 132CoCo-Decision-4Bv1.7 · 10 s rule32.779.756.9$0.0047*
  123. 133APUS-OpenJev-v1-9Bv1.7 · 10 s rule47.268.456.9$0.011*
  124. 134Decider-4Bv1.6.43 drawv1.7 · 10 s rule27.180.756.8$0.0044*
  125. 135typecastlmv1.7 · 10 s rule29.878.556.5$0.0052*
  126. 136Bespoke Nimble 9Bv1.7 · 10 s rule43.156.456.5$0.028*
  127. 137pjev-2bv1.7.0 drawv1.7 · 10 s rule28.863.556.5$0.016*
  128. 138TypeCastLM 1.4.0v1.7 · 10 s rule35.479.356.4$0.0049*
  129. 139RYOTIDE-Qwenv1.6.43 drawv1.7 · 10 s rule25.276.756.3$0.0060*
  130. 140Compass 0.2.0v1.6.43 drawv1.7 · 10 s rule28.972.156.1$0.0085*
  131. 141Standard One 3B SH v4v1.7.0 drawv1.7 · 10 s rule34.448.656.0$0.052*
  132. 142Standard One 3B SHv1.6.43 drawv1.7 · 10 s rule32.848.655.6$0.052*
  133. 143SemIfv1.7 · 10 s rule27.177.655.4$0.0056*
  134. 144ArseneLupin v1.1v1.6.43 drawv1.7 · 10 s rule28.075.255.1$0.0067*
  135. 145local-jev Qwen3.5-4Bv1.7 · 10 s rule24.177.555.0$0.0056*
  136. 146Decision 2Bv1.7 · 10 s rule22.966.454.8$0.013*
  137. 147Jez-large-1.0.0v1.7.0 drawv1.7 · 10 s rule25.683.354.4$0.0036*
  138. 148Nomav1.6.43 drawv1.7 · 10 s rule25.176.554.3$0.0061*
  139. 149spark-s1-4b-v6v1.7 · 10 s rule43.275.054.0$0.0068*
  140. 150Imajev-2Bv1.6.43 drawv1.7 · 10 s rule21.267.354.0$0.012*
  141. 151RYOTIDE-Qwen9v1.6.4 drawv1.7 · 10 s rule27.666.152.5$0.013*
  142. 152Arbiter-4Bv1.6.43 drawv1.7 · 10 s rule27.480.752.3$0.0044*
  143. 153DecisionNode Edge 0.9.2v1.7.0 drawv1.7 · 10 s rule26.886.852.1$0.0028*
  144. 154D1A on Gemma 4 E4Bv1.7.7 drawv1.7 · 10 s rule19.981.251.1$0.0042*
  145. 155ClassOne Qwen 3.5 9Bv1.7 · 10 s rule42.269.050.5$0.011*
  146. 156kev 8Bv1.7 · 10 s rule29.368.950.0$0.011*
  147. 157Liquid AI d1-3Bv1.6.9 drawv1.7 · 10 s rule22.150.649.5$0.044*
  148. 158ZeroEntropy zerank-2v1.7 · 10 s rule15.660.649.4$0.021*
  149. 159JADEv1.7 · 10 s rule47.961.149.2$0.020*
  150. 160open-alternative-jevv1.7 · 10 s rule22.363.349.1$0.017*
  151. 161OpenSourceJevv1.7 · 10 s rule23.283.848.6$0.0035*
  152. 162Open-Jev 2Bv1.7 · 10 s rule21.869.647.7$0.010*
  153. 163EXAONE-4.0-1.2B-JEV v0.3v1.7 · 10 s rule14.857.947.7$0.025*
  154. 164Malkuth-2Bv1.7 · 10 s rule15.983.847.6$0.0035*
  155. 165DecisionNode Edge Mini 0.9.2v1.7.0 drawv1.7 · 10 s rule17.686.847.6$0.0028*
  156. 166Nemotron Diffusion 8Bv1.7 · 10 s rule20.455.547.5$0.030*
  157. 167jiwo 0.8bv1.6.9 drawv1.7 · 10 s rule15.985.047.2$0.0032*
  158. 168decisor-4bv1.6.4 drawv1.7 · 10 s rule39.678.745.7$0.0051*
  159. 169D1A on Gemma 4 E2Bv1.7.7 drawv1.7 · 10 s rule8.984.045.5$0.0034*
  160. 170Qwen3-Reranker-4Bv1.7 · 10 s rule16.360.644.7$0.021*
  161. 171decider-2bv1.7 · 10 s rule20.785.844.6$0.0030*
  162. 172BAAI bge-reranker-v2-m3v1.7 · 10 s rule0.159.344.1$0.023*
  163. 173kev 4Bv1.7 · 10 s rule19.577.944.0$0.0055*
  164. 174Mixedbread mxbai-rerank-base-v2v1.7 · 10 s rule1.160.343.9$0.021*
  165. 175lev-350mv1.7 · 10 s rule4.880.043.3$0.0046*
  166. 176Jez-small-1.0.0v1.7.0 drawv1.7 · 10 s rule4.686.243.2$0.0029*
  167. 177Certo v1v1.7 · 10 s rule0.296.943.2$0.0013*
  168. 178PurpleMIST-Mini-1.0v1.7.0 drawv1.7 · 10 s rule8.486.343.0$0.0029*
  169. 179Chon-1.0-12Bv1.6.43 drawv1.7 · 10 s rule27.364.943.0$0.015*
  170. 180smalljev semantic-v9v1.7 · 10 s rule8.760.843.0$0.020*
  171. 181Vega-Uv1.7.0 drawv1.7 · 10 s rule8.579.442.6$0.0048*
  172. 182jul fastv1.7 · 10 s rule7.569.042.6$0.011*
  173. 183Krino-2Bv1.6.43 drawv1.7 · 10 s rule22.084.242.4$0.0034*
  174. 184Alibaba GTE Reranker ModernBERT-basev1.7 · 10 s rule0.069.341.7$0.011*
  175. 185Jeetu Decider 2Bv1.7.2 drawv1.7 · 10 s rule10.285.941.7$0.0030*
  176. 186watt-flash-0.1v1.7 · 10 s rule3.889.041.6$0.0023
  177. 187Decision Fastv1.7 · 10 s rule6.186.640.6$0.0028*
  178. 188openJev Verdict 1.4v1.7 · 10 s rule5.086.640.4$0.0028*
  179. 189Quyet-1.0-Small-ENv1.7 · 10 s rule6.189.139.5$0.0023*
  180. 190ClassOne Gemma 4 E4Bv1.6.9 drawv1.7 · 10 s rule28.282.138.4$0.0040*
  181. 191ClassOne Qwen 3.5 4Bv1.6.9 drawv1.7 · 10 s rule27.379.538.1$0.0048*
  182. 192OpenDecisionv1.7 · 10 s rule7.479.138.0$0.0050*
  183. 193verdict-smallv1.7 · 10 s rule2.3100.037.6$0.00087*
  184. 194kev 0.6Bv1.7 · 10 s rule7.684.236.9$0.0034*
  185. 195Vega 4Bv1.6.9 drawv1.7 · 10 s rule15.761.736.6$0.019*
  186. 196WaterSheepv1.7 · 10 s rule5.368.735.5$0.011*
  187. 197Raw Phi-4 mini direct logitsv1.7 · 10 s rule13.753.235.3$0.036*
  188. 198Raw Qwen3 4B Instruct 2507 direct logitsv1.7 · 10 s rule34.077.934.6$0.0055*
  189. 199Vega 0.8Bv1.6.9 drawv1.7 · 10 s rule3.168.033.0$0.012*
  190. 200kev 0.5Bv1.7 · 10 s rule4.684.232.9$0.0034*
  191. 201Liquid AI d1-omni-600Mv1.6.9 drawv1.7 · 10 s rule3.373.132.8$0.0079*
  192. 202Quyet-1.0-Smallv1.7 · 10 s rule4.079.632.5$0.0048*
  193. 203Quyet-1.0-Tinyv1.7 · 10 s rule3.788.431.1$0.0024*
  194. 204Open Jev JSON Canvasv1.7 · 10 s rule61.270.730.6$0.0095*
  195. 205Phocinae-Largha-150M-v1v1.6.43 drawv1.7 · 10 s rule3.695.330.2$0.0014*
  196. 206Raw Qwen3 8B direct logitsv1.7 · 10 s rule26.568.830.1$0.011*
  197. 207system-onev1.7 · 10 s rule29.368.329.7$0.011*
  198. 208openJev Verdictv1.7 · 10 s rule4.186.626.6$0.0028*
  199. 209Tacet Sonatav1.7 · 10 s rule2.289.024.2$0.0023*
  200. 210CLM-8Bv1.7 · 10 s rule0.173.720.1$0.0075*
  201. 211ClassOne Qwen 3.5 2Bv1.6.9 drawv1.7 · 10 s rule5.885.819.2$0.0030*
  202. 212Deem 0.8B v1v1.7 · 10 s rule6.586.818.1$0.0027*
  203. 213Laya multilingualv1.7 · 10 s rule0.382.217.0$0.0039*
  204. 214ClassOne Gemma 4 E2Bv1.7 · 10 s rule7.885.815.2$0.0030*
  205. 215Julia-1v1.6.43 drawv1.7 · 10 s rule2.489.79.9$0.0022*
  206. 216Raw Qwen3 1.7B direct logitsv1.7 · 10 s rule5.184.18.3$0.0034*
  207. 217Raw Qwen3 0.6B direct logitsv1.7 · 10 s rule5.684.18.2$0.0034*

Wide coloured bar = Capability Score (0–100). Thin bars: cost above, median latency below; shared log ratio scale 0.25× … 64× Jev (v1.5 reference), ticks at 1× and 2× (cap). Shorter is cheaper or faster. * = est. (estimated cost). # counts ranked Jev-class systems; “–” marks unranked or outside systems. Tap ⓘ or hover a row or thin bar for details.

Cost and latency correlate here (Spearman ρ = 0.21, n = 241).

  • Jev — reference (TypeSafe, closed) (2)
  • Open weights · LLM decoder (205)
  • Open weights · diffusion LM (4)
  • Open weights · encoder / classifier (20)
  • Open weights · reranker (5)
  • Base model control (no decision fine-tune, raw logits) (5)
  • green: ≤ reference
  • amber: ≤ cap (2× reference)
  • red: > cap
Show general-purpose LLMs and other systems outside the limits

Sorted by Capability Score, not numbered. Each row says which limit it misses, measured against Jev ($0.032 per 1,000 decisions, median 0.62 s).

  1. –AJev lm6qv1.7.0 drawv1.7 · 10 s rule71.544.281.4$0.073*Outside: cost 2.2× Jev (v1.5 reference)
  2. –Aleria Judge 1.1.1v1.7.0 drawv1.7 · 10 s rule68.041.176.9$0.092*Outside: cost 2.8× Jev (v1.5 reference)
  3. –OpenJevv1.7 · 10 s rule70.254.975.8$0.032*Outside: latency 3.1× Jev (v1.5 reference)
  4. –SimpleJev Qwen3.8-27Bv1.7 · 10 s rule57.942.574.9$0.083*Outside: cost 2.6× Jev (v1.5 reference), latency 2.3× Jev (v1.5 reference)
  5. –Open-Jev 27B v1.1v1.7 · 10 s rule52.541.768.0$0.088*Outside: cost 2.7× Jev (v1.5 reference), latency 2.9× Jev (v1.5 reference)
  6. –LitJevv1.7 · 10 s rule43.355.766.6$0.030*Outside: latency 5.1× Jev (v1.5 reference)
  7. –Bev / Bonsai 27Bv1.7 · 10 s rule43.855.766.1$0.030*Outside: latency 3.5× Jev (v1.5 reference)
  8. –Open-Jev 9Bv1.7 · 10 s rule43.952.761.8$0.038*Outside: latency 2.3× Jev (v1.5 reference)
  9. –Standard One 8B SHv1.6.9 drawv1.7 · 10 s rule34.543.260.0$0.078*Outside: cost 2.4× Jev (v1.5 reference)
  10. –reflex 4Bv1.7 · 10 s rule30.877.758.9$0.0055*Outside: latency 4.7× Jev (v1.5 reference)
  11. –Standard One 8Bv1.7 · 10 s rule32.843.358.4$0.078*Outside: cost 2.4× Jev (v1.5 reference)
  12. –Nev-2Bv1.6.43 drawv1.6 · re-measure pending20.286.649.9$0.0028*Outside: latency 20.8× Jev (v1.5 reference)
  13. –Laya typed-decisionsv1.7 · 10 s rule5.582.544.2$0.0038*Outside: latency 3.5× Jev (v1.5 reference)
  14. –Qwen3.5-0.8B Decision Modelv1.6 · re-measure pending7.386.142.6$0.0029*Outside: latency 18.3× Jev (v1.5 reference)
  15. –Gutsy 0.8B v0.3v1.6.9 drawv1.7 · 10 s rule8.684.942.2$0.0032*Outside: latency 3.1× Jev (v1.5 reference)
  16. –Reflex Instinct 0.6Bv1.6.43 drawv1.6 · re-measure pending4.086.438.6$0.0028*Outside: latency 6.3× Jev (v1.5 reference)
  17. –Vega-Uv1.6.43 drawv1.7 · 10 s rule5.385.836.1$0.0030*Outside: latency 2.6× Jev (v1.5 reference)
  18. –open-jev-deberta-v3-largev1.7 · 10 s rule2.877.635.4$0.0056*Outside: latency 5.7× Jev (v1.5 reference)
  19. –modernbert-jev v2v1.6.9 drawv1.6 · re-measure pending2.670.434.6$0.0097*Outside: latency 8.7× Jev (v1.5 reference)
  20. –Layav1.7 · 10 s rule1.884.932.5$0.0032*Outside: latency 3.3× Jev (v1.5 reference)
  21. –SimpleJevv1.7 · 10 s rule12.776.931.1$0.0059*Outside: latency 14.9× Jev (v1.5 reference)
  22. –Mirrorv1.7 · 10 s rule0.089.325.8$0.0023*Outside: latency 2.1× Jev (v1.5 reference)

Jev-class = cost per decision at most 2× Jev's (≤ $0.065 per 1,000 decisions) and median latency at most 2× Jev's (≤ 1.23 s), the adjusted p50 — the same median the speed chart plots, not the four-axis Speed score. 219 of 241 systems qualify; the other 22, including the general-purpose LLMs, are listed below the divider in the ranking. The charts below show speed and cost beside Capability Score; the official JevBench Score weighs all four axes.

Capability against cost and speed (open weights)

Jev-class systems are shown by default. Bubble size follows the official JevBench Score; official rank stays unchanged. The five most capable Jev-class systems are labelled.

Capability vs cost

Upper right is best: more capable and cheaper. Cost is USD per 1,000 decisions on a log scale. The dashed line is 2× Jev’s cost.

Capability vs cost: 219 systems. Upper right is best: more capable and cheaper.102030405060708090100$0.0010$0.010$0.10$ per 1,000 decisions (log)Capability↑2× Jev (v1.5 reference)← priciercheaper →1. baselabs mammoth-12. KnowLine 31B-Gen13. Fastino GLiDE 1.24. H2O-Lightning-31B v1.0.75. Perplexity Decider v1.1 27B
219 systems. Tap a bubble for its values.

Capability vs speed

Upper right is best: more capable and faster. Speed here is the median-latency speed — the same adjusted median (p50) latency the Jev-class limit uses — a log scale, so each 20 points is 10× faster (median under the numbers). Self-hosted systems are plotted at the est. incl. network latency (2x measured + 150 ms); their measured loopback p50 (on-GPU, no network; not directly comparable with API latency) is in the tooltip. The dashed line is 2× Jev’s median latency.

Capability vs speed: 219 systems. Upper right is best: more capable and faster.010203040506070809010070≈3.2 s80≈1.0 s90≈316 ms100≈100 msMedian-latency speedCapability↑2× Jev (v1.5 reference)← slowerfaster →1. baselabs mammoth-12. KnowLine 31B-Gen13. Fastino GLiDE 1.24. H2O-Lightning-31B v1.0.75. Perplexity Decider v1.1 27B
219 systems. Tap a bubble for its values.
  • Jev — reference (TypeSafe, closed) (2)
  • Open weights · LLM decoder (205)
  • Open weights · diffusion LM (4)
  • Open weights · encoder / classifier (20)
  • Open weights · reranker (5)
  • Base model control (no decision fine-tune, raw logits) (5)
  • faint = outside Jev-class

Pareto frontier: capability against cost and speed (open weights)

The red line joins the systems nobody beats on both axes at once; blue marks Jev, the reference, on or off the frontier; hover or tap a point for the system that beats it. Shaded: beyond the Jev-class cap.

Speed axis: self-hosted systems are plotted at est. incl. network (2x measured + 150 ms), the value the ranking and cap use; the measured loopback p50 is in each point's tooltip.

Capability vs cost ($ per 1,000 decisions)
Over 2× Jev0102030405060708090100$0.0001$0.0002$0.0005$0.001$0.002$0.005$0.01$0.02$0.05$0.1$ per 1,000 decisions (log scale) — further left is betterCapability ScoreJev 1.13.0 (reference)Jev 1.14.0baselabs mammoth-1Fastino GLiDE 1.2Jeff-1.1-MediumTorchcast Decision 4BXevn-4BChon-1.0-E4BDecisionNode Edge 0.9.2Certo v1verdict-small

Jev (TypeSafe reference)Pareto frontier

9 of 219 systems on the frontier: no other system is both cheaper and more capable.

Capability vs speed (median latency)
Outside the Jev-class cap (2× Jev)0102030405060708090100100 ms200 ms500 ms1.0 s2.0 s5.0 s10 sMedian latency (log scale) — further left is betterCapability ScoreJev 1.13.0 (reference)Jev 1.14.0baselabs mammoth-1Fastino GLiDE 1.2H2O-Lightning-31B v1.0.7StartLux-Decision v1.1 27…Torchcast Decision 4BRespan Span-02Torchcast Decision 4BPlumb-4BJevK5 v0.3Decision 4B v1.1Liquid AI d1-3Bdecider-2bQuyet-1.0-Small-ENQuyet-1.0-SmallQuyet-1.0-Tiny

Jev (TypeSafe reference)Pareto frontier

18 of 219 systems on the frontier: no other system is both faster and more capable.

296 of 296 systems

Model kind

Jev-class

Release

Parameters

No system in this release reports an exact parameter count.

Developer/API price $/1k
Base-model price $/1k
Scored cost $/1k
Alternative price $/1k
p50 latency (s)
p95 latency (s)

JevBench board

JevBench Composite Score (open weights): 235 ranked systems

Official· four axes 0–100, equal-weight harmonic mean · Method notes ↓

Weights:
Adjust weights ↓

View by:Capability ↑

The official order weighs Intelligence, Calibration, Speed and Cost equally. Each button re-sorts the same systems by one axis.

Greener = stronger within its column.

241 of 241 systems, sorted by official rank, #1 first.

  1. 1Torchcast Decision 4B (v1.3)newfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource80.5I 67C 88S 92K 79est.$0.0049
  2. 2H2O-Lightning-4B v1.1newfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource77.9I 60C 90S 92K 79est.$0.0052
  3. 3Wald 26B-A4Bnewfine-tune · 26B-A4B · bf16Base model: google/gemma-4-26B-A4B-itsource77.5I 66C 88S 92K 71est.$0.0093
  4. 4Torchcast Decision 4B (v1.2)newfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource77.0I 60C 85S 91K 79est.$0.0049
  5. 5Fastino GLiDE 1.2newfine-tune · 27B · bf16Base model: undisclosed75.9I 76C 93S 85K 58est.$0.025
  6. 6Bobcat Flash 1.2newfine-tune · 26B / 4B active · FP8 on loadBase model: google/gemma-4-26B-A4B-itsource75.8I 66C 81S 91K 71est.$0.0092
  7. 7H2O-Lightning-31B v1.0.7newfine-tune · 31BBase model: google/gemma-4-31B-itsource75.7I 79C 89S 90K 56est.$0.029
  8. 8StartLux-Decision v1.1 27B Instantnewfine-tune · 27BBase model: undisclosed75.6I 73C 92S 90K 58est.$0.025
  9. 9baselabs mammoth-1newfine-tune · 27B · bf16Base model: Qwen/Qwen3.5-27Bsource75.4I 76C 94S 84K 57est.$0.027
  10. 10Decisio 12B tuned commitnewmerge · 12B · BF16Base model: google/gemma-4-12B-itsource75.2I 63C 89S 90K 67est.$0.013
  11. 11Jeff-1.1-Mediumnewfine-tune · 12B · BF16Base model: google/gemma-4-12B-itsource75.1I 74C 89S 74K 66est.$0.013
  12. 12Philly Decider v2newfine-tune · 4B · BF16Base model: h2oai/h2o-lightning-4bsource75.0I 56C 83S 93K 79est.$0.0050
  13. 13Perplexity Decider v1.1 27Bnewfine-tune · 27BBase model: Qwen/Qwen3.8-27Bsource74.9I 76C 90S 86K 57est.$0.026
  14. 14Chon-1.0-E4Bnewfine-tune · 4BBase model: google/gemma-4-E4B-itsource74.7I 55C 82S 89K 84est.$0.0035
  15. 15vLLM-SR d3-turbonewfine-tune · 27BBase model: undisclosed74.7I 75C 90S 86K 57est.$0.027
  16. 16vLLM-SR d3newfine-tune · 26BBase model: Qwen/Qwen3.8-27Bsource74.6I 73C 91S 87K 57est.$0.027
  17. 17KnowLine 26B-A4B-Gen2.5newfine-tune · 26B-A4BBase model: google/gemma-4-26B-A4B-itsource74.4I 65C 89S 89K 62est.$0.018
  18. 18Torchcast Decision 4B (v1.1)newfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource74.4I 54C 83S 94K 79est.$0.0049
  19. 19KnowLine 35B-A3B-Gen1.5newfine-tune · 35B-A3B · fp8Base model: Qwen/Qwen3.5-35B-A3Bsource74.1I 60C 93S 92K 64est.$0.016
  20. 20basal-omni-1.0-ultranewmerge · 31BBase model: google/gemma-4-31B-itsource74.1I 74C 90S 87K 56est.$0.029
  21. Jev 1.13.0 (reference)APIJev referenceBase model: undisclosed71.4I 64C 91S 91K 55$0.032
  22. Jev 1.14.0 (reference)APInewJev referenceBase model: undisclosed70.2I 60C 90S 91K 55est.$0.032
Show all 241 systems (215 more ranked, 4 more not ranked)
  1. 21Xor 26B-A4Bnewmerge · 26B-A4BBase model: google/gemma-4-26B-A4B-itsource74.0I 57C 85S 91K 72est.$0.0085
  2. 22torchcast-decision-27bnewmerge · 27BBase model: startlux-models/StartLux-Decision-27Bsource 1source 2*174.0I 70C 90S 88K 58est.$0.026
  3. 23Arcbright Elf 1 31Bnewfine-tune · 31B · bf16Base model: google/gemma-4-31B-itsource74.0I 72C 92S 90K 55est.$0.031
  4. 24pplx-decider-v1.2newfine-tune · 27BBase model: Qwen/Qwen3.8-27Bsource73.9I 72C 90S 86K 57est.$0.026
  5. 25Quyet-1.0-Largenewadapter · 31BBase model: google/gemma-4-31B-itsource73.8I 73C 90S 86K 56est.$0.030
  6. 26Nimble-v4-maxnewfine-tune · 27BBase model: Qwen/Qwen3.8-27Bsource73.8I 71C 91S 85K 57est.$0.026
  7. 27decider-12b v2newfine-tune · 12BBase model: google/gemma-4-12B-itsource73.7I 62C 81S 90K 68est.$0.012
  8. 28Xevn-4Bnewfine-tune · 4B · bf16Base model: Qwen/Qwen3.5-4Bsource73.7I 52C 85S 92K 81est.$0.0045
  9. 29Surogate Rune 26B-A4B v3new26B-A4BBase model: google/gemma-4-26B-A4B-itsource*273.4I 56C 91S 87K 71est.$0.0092
  10. 30Jeff-1.0-Largenewfine-tune · 31BBase model: google/gemma-4-31B-itsource73.4I 72C 91S 88K 55est.$0.033
  11. 31Lychee-27Bnewfine-tune · 27B · bf16Base model: Qwen/Qwen3.8-27Bsource73.3I 74C 85S 84K 57est.$0.026
  12. 32decisio v0.8.0 on gemma-4-31B-itnew31B · FP8 on loadBase model: google/gemma-4-31B-itsource73.2I 70C 89S 91K 55est.$0.032
  13. 33decider-12b v1, stock Gemma-4-12B-itnew12BBase model: google/gemma-4-12B-itsource73.2I 60C 83S 90K 68est.$0.012
  14. 34vLLM-SR d3-mininewfine-tune · 4B · bf16Base model: Qwen/Qwen3.5-4Bsource73.2I 51C 87S 93K 79est.$0.0051
  15. 35Sieve-31B-v2newfine-tune · 31B · BF16Base model: google/gemma-4-31B-itsource73.1I 69C 93S 86K 56est.$0.029
  16. 36Vera Largenewfine-tune · 27B · bf16Base model: Qwen/Qwen3.8-27Bsource72.8I 68C 91S 88K 56est.$0.029
  17. 37iota-decide-4b-v0.3newfine-tune · 4B · BF16Base model: Qwen/Qwen3.5-4Bsource72.8I 53C 85S 87K 79est.$0.0051
  18. 38swanOneNVFP4Base model: Mia-AiLab/Qwen3.8-Flash-Next-NVFP4source72.5I 56C 89S 83K 71est.$0.0090
  19. 39Winnow-12B Q8merge · 12B · GGUF Q8_0Base model: google/gemma-4-12B-itsource72.2I 60C 83S 86K 67est.$0.013
  20. 40Bobcat 1.3newfine-tune · 31B · fp8Base model: google/gemma-4-31B-itsource72.2I 68C 88S 89K 55est.$0.031
  21. 41JPT-35B-A3Bnewmerge · 35B-A3BBase model: Qwen/Qwen3.5-35B-A3Bsource72.2I 51C 90S 90K 72est.$0.0083
  22. 42Chon-1.0-12B-Fastnewadapter · 12B · bf16Base model: google/gemma-4-12B-itsource72.0I 60C 80S 84K 69est.$0.011
  23. 43baselabs elephant-1newfine-tune · 9B · bf16Base model: Qwen/Qwen3.5-9Bsource72.0I 56C 85S 90K 68est.$0.012
  24. 44metask-jev-rain-31B-trainednewadapter · 31B · W4A16 INT4Base model: google/gemma-4-31B-itsource72.0I 71C 87S 82K 56est.$0.029
  25. 45deck-31Bnew31B · FP8Base model: google/gemma-4-31B-itsource71.9I 73C 82S 86K 55est.$0.032
  26. 46Cygnet12BBase model: google/gemma-4-12B-itsource71.9I 55C 87S 91K 67est.$0.013
  27. 47metask-jev-rain-12B-trainednewadapter · 12B · W4A16 INT4Base model: google/gemma-4-12B-it-qat-q4_0-unquantizedsource71.7I 66C 87S 72K 65est.$0.014
  28. 48DOPEST 1.0 31Bnewfine-tune · 31B · BF16Base model: google/gemma-4-31B-itsource71.7I 74C 90S 82K 53est.$0.038
  29. 49Eikos-27Badapter · 27BBase model: Qwen/Qwen3.8-27Bsource71.6I 64C 90S 88K 56est.$0.029
  30. 50Jeffan-31b-1710newfine-tune · 31B · bf16Base model: google/gemma-4-31B-itsource71.3I 72C 92S 85K 51est.$0.042
  31. 51Kev 27Bnewfine-tune · 27BBase model: Qwen/Qwen3.8-27Bsource71.3I 64C 94S 83K 56est.$0.029
  32. 52Jev-Omnifine-tuneBase model: google/gemma-4-12B-itsource71.0I 56C 87S 85K 66est.$0.013
  33. 53GEV-26B-Decidenewdistilled · 26B-A4BBase model: google/gemma-4-26B-A4B-itsource70.8I 50C 90S 90K 73est.$0.0082
  34. 54decisio v0.8.0 on gemma-4-12B-itnew12BBase model: google/gemma-4-12B-itsource70.6I 52C 89S 86K 68est.$0.012
  35. 55RSI-Jev v6.1-VL 27Bnewmerge · 27BBase model: Qwen/Qwen3.8-27Bsource70.4I 62C 85S 85K 59est.$0.024
  36. 56Nagi-ENORMOUS 27Bnewadapter · 27BBase model: Qwen/Qwen3.8-27Bsource70.3I 62C 85S 86K 57est.$0.026
  37. 57StartLux-Decision v1.1 27B AutoThinknewfine-tune · 27BBase model: undisclosed70.2I 67C 91S 75K 57est.$0.028
  38. 58metask-jev-rain-31B-trained service v3newadapter · 31B · W4A16 INT4Base model: google/gemma-4-31B-itsource70.2I 73C 87S 73K 55est.$0.031
  39. 59decisio v0.9.0 on gemma-4-12B-itnew12BBase model: google/gemma-4-12B-itsource70.1I 52C 88S 85K 68est.$0.012
  40. 60Bespoke Nimble 9B v3newadapter · 9BBase model: Qwen/Qwen3.5-9Bsource70.0I 57C 77S 89K 66est.$0.014
  41. 61JEV-27Bnewdistilled · 27BBase model: Qwen/Qwen3.8-27Bsource69.9I 57C 91S 86K 59est.$0.024
  42. 62torchcast-decision-12bnewadapter · 12BBase model: google/gemma-4-12B-itsource69.8I 61C 83S 91K 56est.$0.028
  43. 63Messier One v0.4newfine-tune · 4BBase model: undisclosed69.8I 52C 73S 89K 76est.$0.0065
  44. 64shisa-de-1newfine-tune · 26B-A4BBase model: google/gemma-4-26B-A4B-itsource69.7I 50C 81S 93K 71est.$0.0092
  45. 65Cleffine-tuneBase model: Qwen/Qwen3.8-27BsourceWorkers AI price scenario (latency unmeasured): 29.4 (would be #133)69.0I 59C 92S 83K 55est.$0.030
  46. 66Wald 4B v2.1newfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource68.9I 52C 87S 93K 61est.$0.020
  47. 67Hopper 12B trainednewfine-tune · 12BBase model: google/gemma-4-12B-itsource68.8I 52C 85S 84K 67est.$0.013
  48. 68Blink v0.3 26B-A4Bnewfine-tune · 26B-A4B · NVFP4Base model: google/gemma-4-26B-A4B-itsource68.7I 49C 91S 91K 71est.$0.0092
  49. 69Decision 2.0 Vega 27Bnewadapter · 27BBase model: Qwen/Qwen3.8-27Bsource 1source 268.7I 60C 82S 83K 57est.$0.027
  50. 70René-1 31B FP8newfine-tune · 31B · FP8Base model: google/gemma-4-31B-itsource68.6I 62C 90S 87K 51est.$0.042
  51. 71AutoJev-27Bfine-tune · 27BBase model: Qwen/Qwen3.8-27Bsource68.6I 56C 88S 88K 57est.$0.028
  52. 72OpenJev26B-A4BBase model: google/diffusiongemma-26B-A4B-itsource 1source 268.4I 70C 81S 73K 55est.$0.032
  53. 73Jebadiah 27Bnewmerge · 27BBase model: Qwen/Qwen3.8-27Bsource68.3I 57C 83S 85K 57est.$0.026
  54. 74decider chat on Gemma-4-31B-itnew31BBase model: google/gemma-4-31B-itsource68.1I 59C 82S 86K 56est.$0.029
  55. 75NInfer Qwen3.8-Flash-Next mixedLLM decoderBase model: Qwen3.8-Flash-Nextsource67.8I 49C 90S 89K 72est.$0.0087
  56. 76Metask rain 4Bnewfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource67.7I 51C 81S 78K 70est.$0.010
  57. 77FrontiersMind Lumma-fev-27Bnewfine-tune · 27B · bf16Base model: Qwen/Qwen3.8-27Bsource67.6I 55C 89S 85K 55est.$0.032
  58. 78Respan Span-02newLLM decoderBase model: google/gemma-4-26B-A4B-itsource67.1I 61C 78S 86K 54est.$0.035
  59. 79NInfer Qwen3.8-27B NVFP427BBase model: Qwen3.8-27Bsource66.7I 52C 84S 90K 56est.$0.028
  60. 80decider-35b-a3bfine-tune · 35B-A3BBase model: Qwen/Qwen3.5-35B-A3B-Basesource66.5I 49C 84S 91K 73est.$0.0078
  61. 81Diffusion Jevnew26B / 4B activeBase model: google/diffusiongemma-26B-A4B-itsource66.5I 55C 61S 86K 71est.$0.0091
  62. 82SPX-CD-Omninewfine-tune · 12BBase model: google/gemma-4-12B-itsource66.0I 49C 87S 89K 69est.$0.011
  63. 83KnowLine 31B-Gen1newfine-tune · 31B · bf16Base model: google/gemma-4-31B-itsource60.2I 77C 93S 81K 46est.$0.061
  64. 84Amazon Strands Decider 26B-A4Bnewadapter · 26B-A4B · bf16Base model: google/gemma-4-26B-A4B-itsource60.0I 47C 87S 88K 72est.$0.0086
  65. 85Turion Jev 31Bnewfine-tune · 31B · fp8Base model: google/gemma-4-31B-itsource59.6I 71C 91S 86K 47est.$0.060
  66. 86KnowLine 4B-Gen5.5newfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource59.2I 46C 86S 93K 70est.$0.010
  67. 87JevOneadapterBase model: Qwen/Qwen3.6-35B-A3Bsource58.7I 46C 92S 90K 73est.$0.0081
  68. 88Seb-9B-v2newfine-tune · 9B · bf16Base model: Qwen/Qwen3.5-9Bsource57.6I 46C 84S 92K 69est.$0.011
  69. 89pplx-decider-flash-v1.2newfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource56.7I 45C 78S 91K 79est.$0.0050
  70. 90APUS-OpenJev-v1-9Bnewfine-tune · 9BBase model: Qwen/Qwen3.5-9Bsource56.6I 47C 67S 82K 68est.$0.011
  71. 91Aplomb 1newfine-tune · 5.3BBase model: Qwen/Qwen3.5-4Bsource56.0I 45C 89S 89K 72est.$0.0083
  72. 92JADEnewadapter · 27BBase model: Qwen/Qwen3.8-27Bsource53.5I 48C 51S 85K 61est.$0.020
  73. 93Tern-35B-A3Bnewfine-tune · 35B-A3B · bf16Base model: Qwen/Qwen3.6-35B-A3Bsource52.7I 44C 89S 91K 71est.$0.0090
  74. 94AJev lm6qnewadapter · 31B · bf16Base model: google/gemma-4-31B-itsource52.3I 71C 91S 82K 44est.$0.073
  75. 95Arcbright Elf 1 12Bnewfine-tune · 12B · bf16Base model: google/gemma-4-12B-itsource51.0I 44C 86S 92K 68est.$0.012
  76. 96Plumb-4Bfine-tune · 4BBase model: alibiserikbay/JevK5source50.4I 43C 88S 94K 78est.$0.0056
  77. 97Messier One v0.2newfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource47.4I 42C 74S 93K 79$0.0048
  78. 98spark-s1-4b-v6adapter · 4BBase model: Qwen3.5-4Bsource47.1I 43C 65S 86K 75est.$0.0068
  79. 99Bev / Bonsai 27B27B · GGUF PQ2_0Base model: Qwen/Qwen3.8-27Bsource46.6I 44C 89S 73K 56est.$0.030
  80. 100Clef-Flashfine-tuneBase model: Qwen/Qwen3.5-9BsourceWorkers AI price scenario (latency unmeasured): 39.1 (would be #119)46.1I 42C 85S 88K 66est.$0.013
  81. 101Quyet-1.0-Mediumnewadapter · 4BBase model: Qwen/Qwen3.5-4Bsource45.7I 42C 77S 89K 78est.$0.0054
  82. 102Open-Jev 9Badapter · 9BBase model: Qwen/Qwen3.5-9Bsource45.5I 44C 80S 74K 53est.$0.038
  83. 103LitJev27BBase model: Qwen/Qwen3.8-27Bsource44.8I 43C 90S 68K 56est.$0.030
  84. 104Bespoke Nimble 9Badapter · 9BBase model: Qwen3.5-9Bsource44.5I 43C 70S 86K 56est.$0.028
  85. 105SimpleJev Qwen3.8-27Bnew27BBase model: Qwen/Qwen3.8-27Bsource44.5I 58C 92S 76K 42est.$0.083
  86. 106Qwen3.5-9B Jev-like data-mix v2adapter · 9BBase model: Qwen/Qwen3.5-9Bsource44.4I 42C 80S 85K 65est.$0.014
  87. 107levadapter · 4BBase model: Qwen/Qwen3.5-4Bsource44.3I 41C 82S 88K 76est.$0.0061
  88. 108JPT-4Bnewmerge · 4BBase model: Qwen/Qwen3.5-4Bsource44.2I 41C 87S 88K 79est.$0.0051
  89. 109jev-local9BBase model: Qwen/Qwen3.5-9Bsource43.8I 42C 73S 73K 79est.$0.0052
  90. 110Aleria Judge 1.1.1new31B · bf16Base model: undisclosed43.5I 68C 86S 86K 41est.$0.092
  91. 111decider-4b v2adapter · 4BBase model: Qwen/Qwen3.5-4B-Basesource43.1I 40C 90S 90K 79est.$0.0050
  92. 112Jeffan-4B-2010newfine-tune · 4B · bf16Base model: Qwen/Qwen3.5-4Bsource43.0I 40C 88S 93K 79est.$0.0051
  93. 113ClassOne Qwen 3.5 9Bnew9BBase model: Qwen/Qwen3.5-9Bsource43.0I 42C 59S 90K 69est.$0.011
  94. 114Wald 4Bnewfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource42.4I 41C 78S 82K 71est.$0.0096
  95. 115Open-Jev 27B v1.1adapter · 27BBase model: Qwen/Qwen3.8-27Bsource40.5I 52C 83S 72K 42est.$0.088
  96. 116GLiClass v4 9Bnewfine-tune · 9B · BF16Base model: Qwen/Qwen3.5-9Bsource40.2I 40C 87S 86K 71est.$0.0093
  97. 117metask-jev-4b4BBase model: Qwen3.5-4Bsource39.9I 39C 84S 90K 76est.$0.0063
  98. 118Kahn1 4Bnewfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource39.6I 40C 84S 89K 65est.$0.015
  99. 119JevK5 v0.3adapterBase model: Qwen/Qwen3.5-4Bsource39.2I 39C 90S 94K 78est.$0.0056
  100. 120deck-4B v1.0newadapter · FP8Base model: alibiserikbay/JevK5 (built on Qwen/Qwen3.5-4B)source 1source 239.0I 39C 89S 90K 78est.$0.0054
  101. 121Seb-9Bnewfine-tune · 9B · bf16Base model: Qwen/Qwen3.5-9Bsource37.2I 38C 88S 91K 69est.$0.011
  102. 122Decision-4BadapterBase model: Qwen/Qwen3.5-4Bsource37.0I 38C 80S 91K 80est.$0.0045
  103. 123decisor-4bnewfine-tune · 4B · FP8Base model: Qwen/Qwen3.5-4Bsource36.8I 40C 52S 92K 79est.$0.0051
  104. 124janus 4Bnewfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource36.8I 38C 88S 93K 79est.$0.0050
  105. 125Jebadiah 9B v2.1newmerge · 9BBase model: Qwen/Qwen3.5-9Bsource34.7I 37C 90S 91K 69est.$0.011
  106. 126JevK5 v0.2.0adapter · 4BBase model: Qwen3.5-4Bsource33.6I 36C 89S 91K 78est.$0.0056
  107. 127Decision 2.0 Lux 9BnewLLM decoderBase model: vllm-sr/Decision-1.0-Lux-9Bsource31.5I 36C 87S 90K 68est.$0.012
  108. 128Verdin-4Bnewmerge · 4BBase model: Qwen/Qwen3.5-4Bsource31.3I 35C 84S 90K 77est.$0.0060
  109. 129TypeCastLM 1.4.0new3.8BBase model: Qwen/Qwen3.5-4Bsource31.1I 35C 77S 93K 79est.$0.0049
  110. 130blink-4bnewmerge · 4BBase model: Qwen/Qwen3.5-4Bsource30.8I 35C 88S 91K 78est.$0.0054
  111. 131Imajev-9Bnewadapter · 9BBase model: Qwen/Qwen3.5-9Bsource30.4I 37C 88S 79K 50est.$0.045
  112. 132jiwo 4bnewfine-tune · 4.2BBase model: Qwen/Qwen3.5-4Bsource30.3I 35C 86S 92K 79est.$0.0052
  113. 133Imajev-4Badapter · 4BBase model: Qwen/Qwen3.5-4Bsource30.0I 35C 89S 90K 78est.$0.0055
  114. 134Uniphore-4B-Decidernewfine-tune · 4BBase model: undisclosed28.3I 34C 80S 87K 79est.$0.0050
  115. 135Malkuth-4Badapter · 4BBase model: Qwen/Qwen3.5-4B-Basesource28.2I 34C 88S 89K 77est.$0.0056
  116. 136Hopperadapter · 4BBase model: Qwen/Qwen3.5-4Bsource27.7I 34C 89S 91K 77est.$0.0059
  117. 137Decisio v0.9.0 on Qwen3.6-35B-A3Bnewpipeline · 35B-A3B · FP8Base model: undisclosed27.2I 34C 89S 91K 62est.$0.019
  118. 138Clef-omninewfine-tune · 30B-A3BBase model: Qwen/Qwen3-Omni-30B-A3B-Instructsource26.9I 34C 82S 87K 56est.$0.029
  119. 139JEB-35B-A3Bnewfine-tune · 35B-A3BBase model: Qwen/Qwen3.6-35B-A3Bsource26.8I 33C 90S 93K 72est.$0.0084
  120. 140jqv32BBase model: Qwen3-32Bsource25.8I 34C 90S 82K 55est.$0.032
  121. 141CoCo-Decision-4Bnewfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource25.7I 33C 81S 91K 80est.$0.0047
  122. 142Decision 4B v1.2adapter · 4BBase model: Qwen/Qwen3.5-4Bsource25.6I 32C 88S 94K 78est.$0.0056
  123. 143Manchego v2.1adapter · 4BBase model: Qwen/Qwen3.5-4Bsource25.5I 33C 84S 91K 79est.$0.0051
  124. 144Standard One 3B SH v4newfine-tune · 3B · bf16Base model: StandardThinking/StandardOne-3Bsource24.4I 34C 78S 93K 49est.$0.052
  125. 145Raw Qwen3 4B Instruct 2507 direct logitsoriginal · 4BBase model: Qwen/Qwen3-4B-Instruct-2507 (unmodified)source22.6I 34C 35S 91K 78est.$0.0055
  126. 146Standard One 3B SHnewfine-tune · 3BBase model: StandardThinking/StandardOne-3Bsource21.6I 33C 78S 86K 49est.$0.052
  127. 147Jebadiah 9B v2newmerge · 9BBase model: Qwen/Qwen3.5-9Bsource21.3I 31C 83S 91K 69est.$0.011
  128. 148reflex 4Badapter · 4BBase model: Qwen/Qwen3.5-4Bsource21.3I 31C 87S 69K 78est.$0.0055
  129. 149typecastlm4BBase model: Qwen3.5-4Bsource20.5I 30C 83S 93K 79est.$0.0052
  130. 150iota-decide-4bnewfine-tune · 4B · bf16Base model: Qwen/Qwen3.5-4B-Basesource19.7I 29C 85S 87K 78est.$0.0056
  131. 151Decision 4B v1.1adapter · 4BBase model: Qwen/Qwen3.5-4Bsource19.5I 29C 85S 94K 78est.$0.0056
  132. 152Standard One 8B SHnewfine-tune · 8BBase model: StandardThinking/StandardOne-8Bsource18.7I 34C 85S 84K 43est.$0.078
  133. 153kev 8Badapter · 8BBase model: Qwen/Qwen3-8B-Basesource18.5I 29C 71S 88K 69est.$0.011
  134. 154Compass 0.2.0newadapter · 4BBase model: Qwen/Qwen3.5-4Bsource18.5I 29C 83S 83K 72est.$0.0085
  135. 155pjev-2bnewfine-tune · 2.5BBase model: openbmb/MiniCPM5-2Bsource18.1I 29C 84S 93K 64est.$0.016
  136. 156ArseneLupin v1.1newadapter · 4BBase model: Qwen/Qwen3.5-4Bsource17.3I 28C 82S 90K 75est.$0.0067
  137. 157Standard One 8Badapter · 8BBase model: mistralai/Ministral-3-8B-Instruct-2512-BF16source16.9I 33C 84S 92K 43est.$0.078
  138. 158Arbiter-4Bnewfine-tune · 4B · bf16Base model: Qwen/Qwen3.5-4Bsource16.5I 27C 77S 92K 81est.$0.0044
  139. 159Decider-4Bnewfine-tune · 4B · bf16Base model: Qwen/Qwen3.5-4Bsource16.3I 27C 87S 92K 81est.$0.0044
  140. 160SemIf4BBase model: Qwen/Qwen3.5-4Bsource16.1I 27C 84S 90K 78est.$0.0056
  141. 161RYOTIDE-Qwen9newpipeline · 9B · BF16Base model: Qwen/Qwen3.5-9Bsource16.1I 28C 77S 86K 66est.$0.013
  142. 162ClassOne Gemma 4 E4Bnewfine-tune · E4BBase model: google/gemma-4-E4B-itsource16.0I 28C 49S 92K 82est.$0.0040
  143. 163JEV Qwen3.5-9B Base NVFP49B · NVFP4Base model: ig1/Qwen3.5-9B-NVFP4source16.0I 29C 85S 93K 48est.$0.056
  144. 164DecisionNode Edge 0.9.2newfine-tune · 2B · bf16Base model: Qwen/Qwen3.5-2B-Basesource15.9I 27C 77S 93K 87est.$0.0028
  145. 165ClassOne Qwen 3.5 4Bnewfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource14.8I 27C 49S 91K 80est.$0.0048
  146. 166system-one8BBase model: Qwen3-8Bsource14.7I 29C 30S 91K 68est.$0.011
  147. 167Jez-large-1.0.0newfine-tune · E4B · bf16Base model: google/gemma-4-E4B-itsource14.2I 26C 83S 91K 83est.$0.0036
  148. 168Chon-1.0-12Bnewfine-tune · 12BBase model: google/gemma-4-12B-itsource14.2I 27C 59S 65K 65est.$0.015
  149. 169RYOTIDE-Qwennewpipeline · 4BBase model: undisclosed13.5I 25C 87S 89K 77est.$0.0060
  150. 170NomanewLLM decoderBase model: Qwen/Qwen3.5-4B-Basesource13.3I 25C 84S 93K 76est.$0.0061
  151. 171Raw Qwen3 8B direct logitsoriginal · 8BBase model: Qwen/Qwen3-8B-Basesource12.1I 27C 34S 88K 69est.$0.011
  152. 172local-jev Qwen3.5-4Bfine-tune · 4BBase model: Qwen/Qwen3.5-4Bsource12.0I 24C 86S 86K 78est.$0.0056
  153. 173OpenSourceJevadapter · 4B · GGUF Q4_K_MBase model: Qwen3.5-4Bsource10.6I 23C 74S 76K 84est.$0.0035
  154. 174Decision 2Badapter · 2BBase model: openbmb/MiniCPM5-2Bsource10.3I 23C 87S 90K 66est.$0.013
  155. 175open-alternative-jev4BBase model: Qwen3.5-4Bsource9.4I 22C 76S 91K 63est.$0.017
  156. 176Krino-2Bnewadapter · 2BBase model: Qwen/Qwen3.5-2B-Basesource9.2I 22C 63S 91K 84est.$0.0034
  157. 177Liquid AI d1-3Bnewfine-tune · 3BBase model: LiquidAI/LFM2.5-VL-3Bsource8.9I 22C 77S 94K 51est.$0.044
  158. 178Open-Jev 2Badapter · 2BBase model: Qwen/Qwen3.5-2Bsource8.8I 22C 74S 77K 70est.$0.010
  159. 179Imajev-2Bnewadapter · 2BBase model: Qwen/Qwen3.5-2Bsource8.3I 21C 87S 81K 67est.$0.012
  160. 180decider-2badapter · 2BBase model: Qwen/Qwen3.5-2B-Basesource8.0I 21C 69S 94K 86est.$0.0030
  161. 181D1A on Gemma 4 E4Bnewfine-tune · BF16Base model: google/gemma-4-E4Bsource7.4I 20C 82S 92K 81est.$0.0042
  162. 182Nemotron Diffusion 8B8BBase model: nvidia/Nemotron-Labs-Diffusion-8Bsource7.3I 20C 75S 94K 55est.$0.030
  163. Nev-2B (re-measure pending under the 10 s rule)newdistilled · 2B · GGUF Q8_0Base model: Qwen/Qwen3.5-2B-Basesource7.1I 20C 80S 55K 87est.$0.0028
  164. 183kev 4Badapter · 4BBase model: Qwen/Qwen3-4B-Basesource*36.8I 20C 68S 90K 78est.$0.0055
  165. 184DecisionNode Edge Mini 0.9.2newfine-tune · 0.8B · bf16Base model: Qwen/Qwen3.5-0.8B-Basesource5.4I 18C 78S 93K 87est.$0.0028
  166. 185Malkuth-2Badapter · 2BBase model: empero-ai/Qwen3.8-2B-Distillsource4.1I 16C 79S 92K 84est.$0.0035
  167. 186Qwen3-Reranker-4BRerankerBase model: Qwen/Qwen3-4B-Basesource4.1I 16C 73S 81K 61est.$0.021
  168. 187jiwo 0.8bnewfine-tune · 0.75BBase model: Qwen/Qwen3.5-0.8Bsource4.1I 16C 79S 94K 85est.$0.0032
  169. 188ZeroEntropy zerank-2RerankerBase model: Qwen/Qwen3-4Bsource3.7I 16C 83S 82K 61est.$0.021
  170. 189Vega 4Bnewadapter · 4BBase model: undisclosed3.6I 16C 57S 84K 62est.$0.019
  171. 190EXAONE-4.0-1.2B-JEV v0.3new1.2BBase model: LGAI-EXAONE/EXAONE-4.0-1.2Bsource3.3I 15C 81S 94K 58est.$0.025
  172. 191Raw Phi-4 mini direct logitsoriginalBase model: microsoft/Phi-4-mini-instruct (unmodified)source2.5I 14C 57S 92K 53est.$0.036
  173. 192SimpleJev0.8BBase model: Qwen/Qwen3.5-0.8Bsource2.0I 13C 49S 59K 77est.$0.0059
  174. 193Jeetu Decider 2Bnewfine-tune · 2B · bf16Base model: Qwen/Qwen3.5-2B-Basesource1.2I 10C 73S 91K 86est.$0.0030
  175. 194D1A on Gemma 4 E2Bnewfine-tune · BF16Base model: google/gemma-4-E2Bsource0.8I 9C 82S 93K 84est.$0.0034
  176. 195smalljev semantic-v9adapter · 2BBase model: openbmb/MiniCPM5-2B-Basesource0.8I 9C 77S 90K 61est.$0.020
  177. 196Gutsy 0.8B v0.3newfine-tune · 0.8B · GGUF Q8_0Base model: Qwen/Qwen3.5-0.8Bsource0.8I 9C 76S 72K 85est.$0.0032
  178. 197Vega-U (Qwen3.5-4B reader)newadapter · 4BBase model: undisclosed0.7I 8C 77S 89K 79est.$0.0048
  179. 198PurpleMIST-Mini-1.0newfine-tune · 2BBase model: Qwen/Qwen3.5-2B-Basesource0.7I 8C 78S 89K 86est.$0.0029
  180. 199kev 0.6Badapter · 0.6BBase model: Qwen/Qwen3-0.6B-Basesource0.5I 8C 66S 91K 84est.$0.0034
  181. 200jul fastnew2BBase model: openbmb/MiniCPM5-2Bsource0.5I 7C 78S 91K 69est.$0.011
  182. 201OpenDecisionEncoder / classifierBase model: MoritzLaurer/ModernBERT-large-zeroshot-v2.0source0.5I 7C 69S 86K 79est.$0.0050
  183. 202ClassOne Gemma 4 E2BnewE2BBase model: google/gemma-4-E2B-itsource0.5I 8C 23S 93K 86est.$0.0030
  184. Qwen3.5-0.8B Decision Model (re-measure pending under the 10 s rule)fine-tune · 0.8BBase model: Qwen/Qwen3.5-0.8B-Basesource0.5I 7C 78S 56K 86est.$0.0029
  185. 203Deem 0.8B v1LLM decoderBase model: Qwen/Qwen3.5-0.8Bsource0.3I 7C 30S 85K 87est.$0.0027
  186. 204Decision Fastadapter · 0.6BBase model: Qwen/Qwen3-0.6B-Basesource0.3I 6C 75S 91K 87est.$0.0028
  187. 205Quyet-1.0-Small-ENnewfine-tune · 153MBase model: answerdotai/ModernBERT-basesource0.3I 6C 73S 95K 89est.$0.0023
  188. 206ClassOne Qwen 3.5 2Bnewfine-tune · 2BBase model: Qwen/Qwen3.5-2Bsource0.2I 6C 33S 93K 86est.$0.0030
  189. 207Laya typed-decisionsfine-tune · 421MBase model: ModernBERT-largesource0.2I 5C 83S 69K 83est.$0.0038
  190. 208Vega-U (Qwen3.5-0.8B reader)newadapter · 0.8BBase model: undisclosed0.2I 5C 67S 73K 86est.$0.0030
  191. 209WaterSheepnewEncoder / classifierBase model: answerdotai/ModernBERT-basesource0.2I 5C 66S 77K 69est.$0.011
  192. 210openJev Verdict 1.4Encoder / classifierBase model: knowledgator/gliclass-modern-base-v2.0source0.2I 5C 76S 81K 87est.$0.0028
  193. 211Raw Qwen3 0.6B direct logitsoriginal · 0.6BBase model: Qwen/Qwen3-0.6B-Basesource0.2I 6C 11S 93K 84est.$0.0034
  194. 212lev-350madapterBase model: LiquidAI/LFM2.5-350Msource0.2I 5C 82S 94K 80est.$0.0046
  195. 213Raw Qwen3 1.7B direct logitsoriginal · 1.7BBase model: Qwen/Qwen3-1.7B-Basesource0.1I 5C 12S 93K 84est.$0.0034
  196. 214kev 0.5Badapter · 0.5BBase model: Qwen/Qwen2.5-0.5Bsource0.1I 5C 61S 92K 84est.$0.0034
  197. 215Jez-small-1.0.0newfine-tune · 0.8B · fp32Base model: Qwen/Qwen3.5-0.8Bsource0.1I 5C 82S 93K 86est.$0.0029
  198. 216openJev Verdictfine-tuneBase model: knowledgator/gliclass-modern-base-v2.0source 1source 20.1I 4C 49S 78K 87est.$0.0028
  199. 217Quyet-1.0-Smallnewfine-tune · 328MBase model: aisingapore/SEA-LION-ModernBERT-300Msource0.1I 4C 61S 95K 80est.$0.0048
  200. Reflex Instinct 0.6B (re-measure pending under the 10 s rule)newadapter · 0.6BBase model: Qwen/Qwen3-0.6Bsource0.1I 4C 73S 59K 86est.$0.0028
  201. 218watt-flash-0.1newfine-tune · 140MBase model: jhu-clsp/mmBERT-smallsource0.1I 4C 79S 74K 89$0.0023
  202. 219Quyet-1.0-Tinynewdistilled · 183MBase model: jhu-clsp/mmBERT-smallsource0.1I 4C 58S 95K 88est.$0.0024
  203. 220Phocinae-Largha-150M-v1newfine-tune · 144MBase model: undisclosed0.1I 4C 57S 77K 95est.$0.0014
  204. 221Liquid AI d1-omni-600Mnewfine-tune · 600MBase model: LiquidAI/LFM2.5-Encoder-350Msource0.0I 3C 62S 94K 73est.$0.0079
  205. 222Vega 0.8Bnewadapter · 0.8BBase model: Qwen/Qwen3.5-0.8Bsource0.0I 3C 63S 89K 68est.$0.012
  206. 223open-jev-deberta-v3-largeadapterBase model: microsoft/deberta-v3-largesource0.0I 3C 68S 67K 78est.$0.0056
  207. modernbert-jev v2 (re-measure pending under the 10 s rule)newfine-tune · 150M · ONNX INT8Base model: undisclosed0.0I 3C 67S 58K 70est.$0.0097
  208. 224Julia-1newfine-tune · 144MBase model: jhu-clsp/mmBERT-smallsource0.0I 2C 17S 78K 90est.$0.0022
  209. 225verdict-smallfine-tune · 118MBase model: intfloat/multilingual-e5-smallsource0.0I 2C 73S 89K 100est.$0.0009
  210. 226Tacet Sonatanewfine-tune · 144MBase model: jhu-clsp/mmBERT-smallsource0.0I 2C 46S 79K 89est.$0.0023
  211. 227Layafine-tune · 421MBase model: ModernBERT-largesource0.0I 2C 63S 73K 85est.$0.0032
  212. 228Mixedbread mxbai-rerank-base-v2RerankerBase model: Qwen2.5 (size not stated)source 1source 20.0I 1C 87S 92K 60est.$0.021
  213. 229Laya multilingualfine-tune · 322MBase model: mmBERT-basesource0.0I 0C 34S 77K 82est.$0.0039
  214. 230Certo v1Encoder / classifierBase model: ModernBERT-largesource0.0I 0C 86S 92K 97est.$0.0013
  215. 231CLM-8Bfine-tune · 8BBase model: Qwen/Qwen3-8Bsource0.0I 0C 40S 92K 74est.$0.0075
  216. 232BAAI bge-reranker-v2-m3fine-tuneBase model: BAAI/bge-m3source0.0I 0C 88S 94K 59est.$0.023
  217. 233Alibaba GTE Reranker ModernBERT-basefine-tuneBase model: answerdotai/ModernBERT-basesource0.0I 0C 83S 94K 69est.$0.011
  218. 234MirrorEncoder / classifierBase model: undisclosed0.0I 0C 52S 74K 89est.$0.0023
  219. 235Open Jev JSON Canvas26B-A4BBase model: google/diffusiongemma-26B-A4B-itsource0.0I 61C 0S 85K 71est.$0.0095

Base-model notes

  1. *1 torchcast-decision-27b (base model: startlux-models/StartLux-Decision-27B): Direct base is StartLux-Decision-27B (startlux-models, CC-BY-NC-4.0), itself a modified Alibaba Qwen model; its card does not name the Qwen version. Qwen3.8-27B is inferred from the identical text_config.
  2. *2 Surogate Rune 26B-A4B v3 (base model: google/gemma-4-26B-A4B-it): This model also powers our System1 Models s1-pro service.
  3. *3 kev 4B (base model: Qwen/Qwen3-4B-Base): Evaluated Qwen3 variant; later releases use a different base.
Weights:
Adjust weights ↓

Weights are relative: each axis counts in proportion to its slider. The score stays a weighted harmonic mean with the low-axis gates; the gates still apply when an axis sits at 0. Only equal weights give the official JevBench Score and rank.

Score = 4 / (1/I + 1/C + 1/S + 1/K) (each 0–100; × (axis / 50)² for Intelligence, Speed or Cost below 50)

  • Jev — reference (TypeSafe, closed) (2)
  • Open weights · LLM decoder (205)
  • Open weights · diffusion LM (4)
  • Open weights · encoder / classifier (20)
  • Open weights · reranker (5)
  • Base model control (no decision fine-tune, raw logits) (5)
  • Striped bar = same system under the labelled alternative price assumption
  • Shown, not ranked
I, C, S, K = Intelligence, Calibration, Speed, Cost; the est. pill = estimated cost; ann. = announced price; API = the operator's endpoint saw held-out benchmark inputs, without answers; $/1k decisions = US dollars per 1,000 decisions (not heat-shaded). Names link to each project.

Compare two systems

Pick any two. Radars for the score axes, capability by subject topic and use cases, chance-corrected competence per request type on the open and sealed sets, and competence per tier on each set. Further out is better on every spoke; the link keeps the pair.

The four score axes

Radar: the four score axes, two systemsThe four score axes, Torchcast Decision 4B vs H2O-Lightning-4B v1.1. Intelligence: 67.1 vs 60.0; Calibration: 87.8 vs 90.0; Speed: 92.4 vs 92.4; Cost: 79.4 vs 78.6.50100Intelligence: Torchcast Decision 4B: 67.1; H2O-Lightning-4B v1.1: 60.0Intelligence67.1 · 60.0Calibration: Torchcast Decision 4B: 87.8; H2O-Lightning-4B v1.1: 90.0Calibration87.8 · 90.0Speed: Torchcast Decision 4B: 92.4; H2O-Lightning-4B v1.1: 92.4Speed92.4 · 92.4Cost: Torchcast Decision 4B: 79.4; H2O-Lightning-4B v1.1: 78.6Cost79.4 · 78.6
0–100, the values in the table. An axis a system has no published value for is left as a gap (it counts as 0 in the composite).

Capability by subject topic

Torchcast Decision 4B: Subject topics and use cases come from this draw's own item labels, added after its release with the same labelling recipe as the earlier draws (hand check 71/75).

Radar: capability by subject topic, two systemsCapability by subject topic, Torchcast Decision 4B vs H2O-Lightning-4B v1.1. Rules, policy & law: 73.0 vs 55.0; Coding & software: 70.0 vs 59.9; Math & numbers: 41.6 vs 25.9; Finance & commerce: 78.0 vs 59.5; Support & operations: 76.5 vs 60.2; Everyday language: 45.2 vs 76.6; Safety & security: 72.8 vs 66.5. Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.−1000100Rules, policy & law: Torchcast Decision 4B: 73.0; H2O-Lightning-4B v1.1: 55.0 — Rules, policy & law: applying written rules: company policies, contracts, regulations, eligibility. 1616 items (151 open, 1465 sealed).Rules, policy& law73.0 · 55.0Coding & software: Torchcast Decision 4B: 70.0; H2O-Lightning-4B v1.1: 59.9 — Coding & software: code, SQL, repositories, developer tools and IT systems. 382 items (62 open, 320 sealed).Coding &software70.0 · 59.9Math & numbers: Torchcast Decision 4B: 41.6; H2O-Lightning-4B v1.1: 25.9 — Math & numbers: a calculation decides the answer: arithmetic, word problems, probability, dates, units. 370 items (37 open, 333 sealed).Math &numbers41.6 · 25.9Finance & commerce: Torchcast Decision 4B: 78.0; H2O-Lightning-4B v1.1: 59.5 — Finance & commerce: money: payments, refunds, invoices, orders, expenses, insurance payouts. 223 items (12 open, 211 sealed).Finance &commerce78.0 · 59.5Support & operations: Torchcast Decision 4B: 76.5; H2O-Lightning-4B v1.1: 60.2 — Support & operations: support tickets, incidents, logistics, scheduling desks and routing work to a team. 160 items (21 open, 139 sealed).Support &operations76.5 · 60.2Everyday language: Torchcast Decision 4B: 45.2; H2O-Lightning-4B v1.1: 76.6 — Everyday language: short everyday messages: intents, assistant requests, reading a detail out of a text. 145 items (11 open, 134 sealed).Everydaylanguage45.2 · 76.6Safety & security: Torchcast Decision 4B: 72.8; H2O-Lightning-4B v1.1: 66.5 — Safety & security: untrusted or injected instructions, fraud, moderation, access and security triage. 109 items (6 open, 103 sealed).Safety &security72.8 · 66.5

Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.

Measured subject-topic competence on this release’s item set. Chance-corrected competence per category (0 = at or below chance; negative averages are reported as 0, 100 = perfect), Torchcast Decision 4B: S+P (draw v1.6-regular-20261012); H2O-Lightning-4B v1.1: S+P+L1+L2+L3 items pooled; only categories with at least 30 items are spokes, smaller ones are listed below. Hover a category for its definition and item count.
What each category means · items per category
  • Rules, policy & law — applying written rules: company policies, contracts, regulations, eligibility. 1616 items (151 open / 1465 sealed)
  • Coding & software — code, SQL, repositories, developer tools and IT systems. 382 items (62 open / 320 sealed)
  • Math & numbers — a calculation decides the answer: arithmetic, word problems, probability, dates, units. 370 items (37 open / 333 sealed)
  • Finance & commerce — money: payments, refunds, invoices, orders, expenses, insurance payouts. 223 items (12 open / 211 sealed)
  • Support & operations — support tickets, incidents, logistics, scheduling desks and routing work to a team. 160 items (21 open / 139 sealed)
  • Everyday language — short everyday messages: intents, assistant requests, reading a detail out of a text. 145 items (11 open / 134 sealed)
  • Safety & security — untrusted or injected instructions, fraud, moderation, access and security triage. 109 items (6 open / 103 sealed)

Use cases (TypeSafe categories)

Torchcast Decision 4B: Subject topics and use cases come from this draw's own item labels, added after its release with the same labelling recipe as the earlier draws (hand check 71/75).

Radar: use cases (typesafe categories), two systemsUse cases (TypeSafe categories), Torchcast Decision 4B vs H2O-Lightning-4B v1.1. Model routing: 76.2 vs 67.5; Legal & compliance: 74.8 vs 65.5; Customer support: 65.6 vs 60.5; Other: 51.1 vs 39.0; E-commerce: 66.6 vs 58.6; Insurance claims: n=27 vs 60.6; Risk assessment: 54.7 vs 59.2; Financial crime: 36.1 vs 35.4; Feature extraction: n=19 vs 47.6; Lead generation: 66.9 vs 62.7; Recruiting: n=19 vs 21.7; Knowledge graphs: n=22 vs 34.0; LLM guardrails: n=27 vs 66.5; Moderation: 74.9 vs 55.9; Code linting: n=24 vs 27.8; Search & retrieval: n=24 vs 70.6; Science: n=23 vs 38.4; Advertising: n=22 vs 61.4; Gaming: n=23 vs 18.3; Demand forecasting: n=17 vs 21.2. Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.−1000100Model routing: Torchcast Decision 4B: 76.2; H2O-Lightning-4B v1.1: 67.5 — Model routing: deciding which LLM or model tier should handle a prompt; classifying a request's intent, domain, difficulty or risk to pick a model. 524 items (86 open, 438 sealed).Modelrouting76.2 · 67.5Legal & compliance: Torchcast Decision 4B: 74.8; H2O-Lightning-4B v1.1: 65.5 — Legal and compliance: contracts, policies, regulations, eligibility rules, checking documents against written requirements. 515 items (74 open, 441 sealed).Legal &compliance74.8 · 65.5Customer support: Torchcast Decision 4B: 65.6; H2O-Lightning-4B v1.1: 60.5 — Customer support: classifying tickets and customer intents, urgency, refunds, routing cases to a team, checking support replies. 240 items (37 open, 203 sealed).Customersupport65.6 · 60.5Other: Torchcast Decision 4B: 51.1; H2O-Lightning-4B v1.1: 39.0 — Other: none of the above: general reasoning, arithmetic or date puzzles, everyday assistant requests without a business workflow. 174 items (24 open, 150 sealed).Other51.1 · 39.0E-commerce: Torchcast Decision 4B: 66.6; H2O-Lightning-4B v1.1: 58.6 — E-commerce marketplaces: product listings, product attributes, orders and deliveries, seller catalog policy. 119 items (9 open, 110 sealed).E-commerce66.6 · 58.6Insurance claims: Torchcast Decision 4B: n=27 — too few completed items to plot; H2O-Lightning-4B v1.1: 60.6 — Insurance claims: first-notice-of-loss reports, claim complexity, missing information, claim triage. 119 items (9 open, 110 sealed).Insuranceclaimsn=27 · 60.6Risk assessment: Torchcast Decision 4B: 54.7; H2O-Lightning-4B v1.1: 59.2 — Risk assessment: estimating probabilities and severity of risks, incidents, vendor or underwriting risk. 118 items (8 open, 110 sealed).Riskassessment54.7 · 59.2Financial crime: Torchcast Decision 4B: 36.1; H2O-Lightning-4B v1.1: 35.4 — Financial crime: suspicious transactions, KYC, fraud alerts, entity matching for investigations. 101 items (9 open, 92 sealed).Financialcrime36.1 · 35.4Feature extraction: Torchcast Decision 4B: n=19 — too few completed items to plot; H2O-Lightning-4B v1.1: 47.6 — Feature extraction for predictive modeling: turning natural-language data into probabilistic features or estimates for a downstream prediction. 99 items (5 open, 94 sealed).Featureextractionn=19 · 47.6Lead generation: Torchcast Decision 4B: 66.9; H2O-Lightning-4B v1.1: 62.7 — Lead generation: matching companies or inbound messages to an ideal customer profile, buyer intent, prioritising and routing sales leads. 95 items (4 open, 91 sealed).Leadgeneration66.9 · 62.7Recruiting: Torchcast Decision 4B: n=19 — too few completed items to plot; H2O-Lightning-4B v1.1: 21.7 — Recruiting: resumes, applications, interview feedback, matching candidates to roles. 94 items (1 open, 93 sealed).Recruitingn=19 · 21.7Knowledge graphs: Torchcast Decision 4B: n=22 — too few completed items to plot; H2O-Lightning-4B v1.1: 34.0 — Graphs and knowledge graphs: entity types, relationships between records, contradictions between facts, multi-step lookups across linked facts. 94 items (3 open, 91 sealed).Knowledgegraphsn=22 · 34.0LLM guardrails: Torchcast Decision 4B: n=27 — too few completed items to plot; H2O-Lightning-4B v1.1: 66.5 — LLM guardrails: checking LLM inputs, outputs and tool calls: prompt injection, jailbreaks, policy violations, tool-call errors, answer quality. 93 items (5 open, 88 sealed).LLMguardrailsn=27 · 66.5Moderation: Torchcast Decision 4B: 74.9; H2O-Lightning-4B v1.1: 55.9 — Moderation and trust and safety: user content moderation: toxicity, harassment, spam, unsafe advice, personal data exposure. 93 items (7 open, 86 sealed).Moderation74.9 · 55.9Code linting: Torchcast Decision 4B: n=24 — too few completed items to plot; H2O-Lightning-4B v1.1: 27.8 — Semantic code linting: checking code or writing against conventions and guidelines, as in CI review of code. 90 items (3 open, 87 sealed).Code lintingn=24 · 27.8Search & retrieval: Torchcast Decision 4B: n=24 — too few completed items to plot; H2O-Lightning-4B v1.1: 70.6 — Search and retrieval: scoring query-to-candidate relevance, reranking results, selecting useful context for RAG. 90 items (5 open, 85 sealed).Search &retrievaln=24 · 70.6Science: Torchcast Decision 4B: n=23 — too few completed items to plot; H2O-Lightning-4B v1.1: 38.4 — Scientific discovery: screening papers, labelling research passages or survey answers, checking that citations support claims, methodology checks. 90 items (2 open, 88 sealed).Sciencen=23 · 38.4Advertising: Torchcast Decision 4B: n=22 — too few completed items to plot; H2O-Lightning-4B v1.1: 61.4 — Advertising: ad creatives, campaign copy, brand safety, prohibited claims in ads. 88 items (3 open, 85 sealed).Advertisingn=22 · 61.4Gaming: Torchcast Decision 4B: n=23 — too few completed items to plot; H2O-Lightning-4B v1.1: 18.3 — Gaming: player reports, in-game chat, game support. 86 items (3 open, 83 sealed).Gamingn=23 · 18.3Demand forecasting: Torchcast Decision 4B: n=17 — too few completed items to plot; H2O-Lightning-4B v1.1: 21.2 — Demand forecasting: purchase intent, product interest, demand and supply signals for forecasting. 83 items (3 open, 80 sealed).Demandforecastingn=17 · 21.2

Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.

Measured application-category competence on this release’s item set. Chance-corrected competence per category (0 = at or below chance; negative averages are reported as 0, 100 = perfect), Torchcast Decision 4B: S+P (draw v1.6-regular-20261012); H2O-Lightning-4B v1.1: S+P+L1+L2+L3 items pooled; only categories with at least 30 items are spokes, smaller ones are listed below. Hover a category for its definition and item count.A (Torchcast Decision 4B) is measured on 9 of 20 spokes only, so it is drawn as points, not a shape — measured on fewer items; full radar after the next re-measure.Gaps, not zeros: Open spokes (n/a or n=…) have fewer than 30 answered items for that system or no published value.

Low sample, n < 30 — indicative only

Category (items)A: Torchcast Decision 4BB: H2O-Lightning-4B v1.1
Insurance claims (119; a system answered fewer than 30)71.5 n=2760.6 n=119
Feature extraction for predictive modeling (99; a system answered fewer than 30)40.3 n=1947.6 n=99
Recruiting (94; a system answered fewer than 30)82.9 n=1921.7 n=94
Graphs and knowledge graphs (94; a system answered fewer than 30)52.2 n=2234.0 n=94
LLM guardrails (93; a system answered fewer than 30)59.1 n=2766.5 n=93
Semantic code linting (90; a system answered fewer than 30)24.3 n=2427.8 n=90
Search and retrieval (90; a system answered fewer than 30)97.6 n=2470.6 n=90
Scientific discovery (90; a system answered fewer than 30)58.6 n=2338.4 n=90
Advertising (88; a system answered fewer than 30)52.1 n=2261.4 n=88
Gaming (86; a system answered fewer than 30)50.6 n=2318.3 n=86
Demand forecasting (83; a system answered fewer than 30)23.9 n=1721.2 n=83

Not drawn on the radar: with so few items a single answer moves a category score by several points, so these values are noise-prone. These values become spokes once the item pool reaches 30 items per category.

What each category means · items per category
  • Model routing — deciding which LLM or model tier should handle a prompt; classifying a request's intent, domain, difficulty or risk to pick a model. 524 items (86 open / 438 sealed)
  • Legal and compliance — contracts, policies, regulations, eligibility rules, checking documents against written requirements. 515 items (74 open / 441 sealed)
  • Customer support — classifying tickets and customer intents, urgency, refunds, routing cases to a team, checking support replies. 240 items (37 open / 203 sealed)
  • Other — none of the above: general reasoning, arithmetic or date puzzles, everyday assistant requests without a business workflow. 174 items (24 open / 150 sealed)
  • E-commerce marketplaces — product listings, product attributes, orders and deliveries, seller catalog policy. 119 items (9 open / 110 sealed)
  • Insurance claims — first-notice-of-loss reports, claim complexity, missing information, claim triage. 119 items (9 open / 110 sealed)
  • Risk assessment — estimating probabilities and severity of risks, incidents, vendor or underwriting risk. 118 items (8 open / 110 sealed)
  • Financial crime — suspicious transactions, KYC, fraud alerts, entity matching for investigations. 101 items (9 open / 92 sealed)
  • Feature extraction for predictive modeling — turning natural-language data into probabilistic features or estimates for a downstream prediction. 99 items (5 open / 94 sealed)
  • Lead generation — matching companies or inbound messages to an ideal customer profile, buyer intent, prioritising and routing sales leads. 95 items (4 open / 91 sealed)
  • Recruiting — resumes, applications, interview feedback, matching candidates to roles. 94 items (1 open / 93 sealed)
  • Graphs and knowledge graphs — entity types, relationships between records, contradictions between facts, multi-step lookups across linked facts. 94 items (3 open / 91 sealed)
  • LLM guardrails — checking LLM inputs, outputs and tool calls: prompt injection, jailbreaks, policy violations, tool-call errors, answer quality. 93 items (5 open / 88 sealed)
  • Moderation and trust and safety — user content moderation: toxicity, harassment, spam, unsafe advice, personal data exposure. 93 items (7 open / 86 sealed)
  • Semantic code linting — checking code or writing against conventions and guidelines, as in CI review of code. 90 items (3 open / 87 sealed)
  • Search and retrieval — scoring query-to-candidate relevance, reranking results, selecting useful context for RAG. 90 items (5 open / 85 sealed)
  • Scientific discovery — screening papers, labelling research passages or survey answers, checking that citations support claims, methodology checks. 90 items (2 open / 88 sealed)
  • Advertising — ad creatives, campaign copy, brand safety, prohibited claims in ads. 88 items (3 open / 85 sealed)
  • Gaming — player reports, in-game chat, game support. 86 items (3 open / 83 sealed)
  • Demand forecasting — purchase intent, product interest, demand and supply signals for forecasting. 83 items (3 open / 80 sealed)

Competence per request type, open / sealed

Radar: competence per request type, open / sealed, two systemsCompetence per request type, open / sealed, Torchcast Decision 4B vs H2O-Lightning-4B v1.1. Choice · open: 81.7 vs 75.2; Choice · sealed: 73.0 vs 64.7; Noul · open: 80.1 vs 60.9; Noul · sealed: 70.9 vs 68.1; Score · open: 47.5 vs 47.0; Score · sealed: 49.8 vs 44.3. Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.−1000100Choice · open: Torchcast Decision 4B: 81.7; H2O-Lightning-4B v1.1: 75.2Choice · open81.7 · 75.2Choice · sealed: Torchcast Decision 4B: 73.0; H2O-Lightning-4B v1.1: 64.7Choice ·sealed73.0 · 64.7Noul · open: Torchcast Decision 4B: 80.1; H2O-Lightning-4B v1.1: 60.9Noul · open80.1 · 60.9Noul · sealed: Torchcast Decision 4B: 70.9; H2O-Lightning-4B v1.1: 68.1Noul · sealed70.9 · 68.1Score · open: Torchcast Decision 4B: 47.5; H2O-Lightning-4B v1.1: 47.0Score · open47.5 · 47.0Score · sealed: Torchcast Decision 4B: 49.8; H2O-Lightning-4B v1.1: 44.3Score ·sealed49.8 · 44.3

Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.

Chance-corrected competence for Choice, Noul and Score on the 300 open and 1200 sealed decisions.

Competence per tier — open set

Radar: competence per tier — open set, two systemsCompetence per tier — open set, Torchcast Decision 4B vs H2O-Lightning-4B v1.1. Easy: 86.2 vs 88.2; Standard: 68.5 vs 57.6; Judge: 65.2 vs 61.0; Hard: 75.7 vs 63.7. Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.−1000100Easy: Torchcast Decision 4B: 86.2; H2O-Lightning-4B v1.1: 88.2Easy86.2 · 88.2Standard: Torchcast Decision 4B: 68.5; H2O-Lightning-4B v1.1: 57.6Standard68.5 · 57.6Judge: Torchcast Decision 4B: 65.2; H2O-Lightning-4B v1.1: 61.0Judge65.2 · 61.0Hard: Torchcast Decision 4B: 75.7; H2O-Lightning-4B v1.1: 63.7Hard75.7 · 63.7

Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.

Per-tier competence, the three request types pooled by their published decision counts.

Competence per tier — sealed set

Radar: competence per tier — sealed set, two systemsCompetence per tier — sealed set, Torchcast Decision 4B vs H2O-Lightning-4B v1.1. Easy: 77.4 vs 74.4; Standard: 70.8 vs 62.6; Judge: 65.3 vs 63.2; Hard: 62.1 vs 54.3. Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.−1000100Easy: Torchcast Decision 4B: 77.4; H2O-Lightning-4B v1.1: 74.4Easy77.4 · 74.4Standard: Torchcast Decision 4B: 70.8; H2O-Lightning-4B v1.1: 62.6Standard70.8 · 62.6Judge: Torchcast Decision 4B: 65.3; H2O-Lightning-4B v1.1: 63.2Judge65.3 · 63.2Hard: Torchcast Decision 4B: 62.1; H2O-Lightning-4B v1.1: 54.3Hard62.1 · 54.3

Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.

Per-tier competence on the sealed decisions, types pooled the same way; item text stays private.
How the categories were made

O1S chance-corrected competence: equal mean across the request types present, clipped to 0–100. Zero means the equal-weight mean across present request types was at or below baseline; negative means are clipped. It does not mean every type scored zero. Choice uses a random-option baseline; Noul uses a fixed 50% accuracy baseline (abstentions count as wrong); Score uses the midpoint-guess error (random-guess fallback when every gold is at the midpoint). Compare models within a category; row pools and supported types can differ.

Categories use ruled subject-topic labels and authoring use-case precedence across the answered pool union, with stable item identities counted once.

  • Official O1S; recorded failures count where supported; no missing response is imputed; stable identity dedupe parent first.
  • S-based category cells pool recorded supported responses from S+P and the answered L1/L2/L3 supplements. API overlay category cells keep their separately recorded A4/A5+P union. Headline and per-type/tier scores are unchanged.
All values as a table
SpokeA: Torchcast Decision 4BB: H2O-Lightning-4B v1.1
The four score axes
Intelligence67.160.0
Calibration87.890.0
Speed92.492.4
Cost79.478.6
Capability by subject topic
Rules, policy & law73.055.0
Coding & software70.059.9
Math & numbers41.625.9
Finance & commerce78.059.5
Support & operations76.560.2
Everyday language45.276.6
Safety & security72.866.5
Use cases (TypeSafe categories)
Model routing76.267.5
Legal & compliance74.865.5
Customer support65.660.5
Other51.139.0
E-commerce66.658.6
Insurance claimsn=2760.6
Risk assessment54.759.2
Financial crime36.135.4
Feature extractionn=1947.6
Lead generation66.962.7
Recruitingn=1921.7
Knowledge graphsn=2234.0
LLM guardrailsn=2766.5
Moderation74.955.9
Code lintingn=2427.8
Search & retrievaln=2470.6
Sciencen=2338.4
Advertisingn=2261.4
Gamingn=2318.3
Demand forecastingn=1721.2
Competence per request type, open / sealed
Choice · open81.775.2
Choice · sealed73.064.7
Noul · open80.160.9
Noul · sealed70.968.1
Score · open47.547.0
Score · sealed49.844.3
Competence per tier — open set
Easy86.288.2
Standard68.557.6
Judge65.261.0
Hard75.763.7
Competence per tier — sealed set
Easy77.474.4
Standard70.862.6
Judge65.363.2
Hard62.154.3

Category radars count each answered item once from the pools named under each radar. API overlay rows use A4+P+L1+L2+L3; A5+P+L1+L2+L3; S+P+L1+L2+L3. Raw and unequated; cells under 15 answered items are omitted. Per-type and tier radars retain each row's original measurement pools: A4/A5 rows have 300 open plus 300 sealed items; full-set rows have S 1,200 plus P 300.

Languages

Raw chance-corrected competence per item language (0 = chance, 100 = perfect; can be negative), from each system's own measured items: Self-hosted rows use S 1,200 + P 300 + L1 (354 items) + L2 (33 items) + L3 (1,118 items), where answered. API rows re-run on A4/A5 use their 300/300-item sealed subset instead of S. Every row’s tag lists the pools it actually answered. L3 is a sealed language supplement (1,118 items), drawn 2026-10-07. Every language has at least 60 items in P ∪ L1 ∪ L2 ∪ L3. Header counts show the full S + P + L1 + L2 + L3 pool. Headline scores are unchanged; these raw cells are unequated, scored for language/category views only and outside the Composite. L3 items were written natively by Claude Sonnet 5.5, each solved blind and language-checked by GPT-6.1 Sol, with gold kept only when both agree or a second review confirms; no gold comes from Jev or any measured API. L3 is API-facing by design and is excluded from future headline draws. Rows with unfinished runs retain their actual coverage tags. L3 items were reviewed by an OpenAI model, so the L3-based language and category cells of OpenAI rows (OpenAI Decisions, GPT-6 Luna, GPT-5.6 Luna) carry that exposure; headline scores do not use L3. C1 adds English items for thin radar categories (everyday language, safety and the other use case). L4 adds 20 new Hindi items for Fastino GLiNER 2.5 Decide only. They were authored with OpenAI and independently solved and language-reviewed with Anthropic before sealing. Seventeen L4 requests succeeded; three HTTP 400 failures remain in the scored observations and do not count as completed coverage. L4 is outside headline and category scores and the common header counts. L5 adds 60 new items for OpenJev DeBERTa v3 Large only: 20 each in Arabic, Hindi and Greek. They were authored with OpenAI and independently solved and language-reviewed with Anthropic before sealing, then checked against the original 512-token native context before keeper selection. All 60 native requests succeeded. Previous failures remain scored and do not count as completed coverage. L5 is outside headline and category scores and the common header counts. Cells under 15 completed supported responses are left empty; recorded input refusals count toward coverage. A dagger (†) marks fewer than 30 completed responses. Competence retains all supported scored observations, including failures and refusals. The mixed-language group is listed first, then English (1,306 items); the other 21 languages and the mixed-language group share 1699 items.

355 systems

JevBench v1.6.1 competence by item language and system
781306959088848383808077757575737372727171706965
H2O-Lightning-4B v1.4.1self-hosted · not measured86687575685945648179696976647570686087†86817152†
Torchcast Decision 4B (v1.3)self-hosted · S+P (draw v1.6-regular-20261012)—667275†7168—82†—62†———————70†27†———66†
Torchcast Decision 31B v1.0self-hosted · not measured90908689929890998991959097929387779294†949410092†
WeMM-JEVself-hosted · not measured90838884808572898283828894809389697493†88969075†
Arcbright Elf 1 12B (r2)self-hosted · not measured77697269736839847775757383657478626898†67836363†
z3-mini-soup-previewself-hosted · not measured89617771525834607370675563427269584875†67715655†
Jeff-1.1-Qwen-4Bself-hosted · not measured93706775658766728476636684747878504795†87836662†
iota-decide-4b-v0.4self-hosted · not measured84617372606745576466576682517475576787†73855864†
Decisio 12B tuned commit-thinkself-hosted · not measured83686767637457768171747378657271606391†78806352†
H2O-Lightning-4B v1.1self-hosted · S+P+L1+L2+L38563555960544938545841373446264946394855473123
VisionJEV 1.0.0self-hosted · not measured86818783808467877590748989748789577392†86928885†
Wald 26B-A4Bself-hosted · not measured—646568†7472—60†—36†———————58†29†———78†
Torchcast Decision 4B (v1.2)self-hosted · S+P (draw v1.6-regular-20261012)—606568†7678—71†—63†———————39†9†———74†
Fastino GLiDE 1.2self-hosted · not measured—777478†7991—67†—52†———————70†49†———94†
Bobcat Flash 1.2self-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.8271454738543333514143373928264737485044582927
H2O-Lightning-31B v1.0.7self-hosted · S+P (draw v1.6-regular-20261012)—816675†8391—78†—65†———————75†28†———77†
StartLux-Decision v1.1 27B Instantself-hosted · S+P (draw v1.6-regular-20261012)—727886†7686—78†—51†———————82†61†———78†
Decision-King v1.1self-hosted · not measured95748385779377848683648680749177808188†81887583†
Piccolo 4Bself-hosted · not measured85607470516542586570616471587867584477†76716863†
baselabs mammoth-1self-hosted · not measured—777683†6590—65†—62†———————79†76†———61†
Decisio 12B tuned commitself-hosted · not measured—607066†6183—55†—52†———————56†39†———71†
Jeff-1.1-Mediumself-hosted · not measured—758276†8894—85†—29†———————88†52†———78†
Philly Decider v2self-hosted · not measured—546349†6670—72†—50†———————29†0†———55†
Perplexity Decider v1.1 27Bself-hosted · S+P+L1+L2+L38579505862634352705148425721386333676442643957
Chon-1.0-E4Bself-hosted · not measured—506875†6883—71†—45†———————61†25†———52†
vLLM-SR d3-turboself-hosted · S+P (draw v1.6-regular-20261012)—757479†7683—78†—37†———————57†78†———72†
vLLM-SR d3self-hosted · S+P (draw v1.6-regular-20261012)—747379†7986—73†—39†———————51†69†———55†
Jeff-1.1-Largeself-hosted · not measured96888582808683989890788192809189788989†96838795†
KnowLine 26B-A4B-Gen2.5self-hosted · not measured—677548†6579—79†—19†———————57†49†———50†
Torchcast Decision 4B (v1.1)self-hosted · S+P (draw v1.6-regular-20261012)—546265†7261—80†—46†———————33†0†———59†
oenogarumself-hosted · not measured927774†6980816588†62†84†69†787366†93767674†76†92†946768†
Piccolo 9Bself-hosted · not measured90647561747938716776657565617679665679†80766368†
KnowLine 35B-A3B-Gen1.5self-hosted · not measured—617159†5092—75†—15†———————42†44†———44†
basal-omni-1.0-ultraself-hosted · S+P (draw v1.6-regular-20261012)—777670†6586—62†—58†———————71†39†———78†
Umpire 35B-A3Bself-hosted · not measured78577462557632767562537061517262474388†75756169†
Xor 26B-A4Bself-hosted · S+P+L1+L2+L38061464048565136524434454332335040685050504246
allec-oavgself-hosted · not measured957578†6178667488†68†75†69†737059†88736970†76†87†948268†
torchcast-decision-27bself-hosted · S+P+L1+L2+L38174495851454340675242534723215038576139513544
Arcbright Elf 1 31Bself-hosted · not measured—726771†7486—59†—50†———————73†24†———77†
pplx-decider-v1.2self-hosted · S+P (draw v1.6-regular-20261012)—737877†7887—73†—35†———————43†48†———61†
MetaDecision-31B-Rainself-hosted · not measured9890949690927610093949178100929690788895†1009410079†
Quyet-1.0-Largeself-hosted · S+P+L1+L2+L38476475756665440665543525943296459676746554564
Nimble-v4-maxself-hosted · not measured—746965†7983—72†—31†———————52†69†———55†
decider-12b v2self-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.8765515861404955614954435546435759675856405836
Xevn-4Bself-hosted · not measured—496372†4472—77†—56†———————37†25†———70†
cribl-decision-1.0-27Bself-hosted · not measured90646864627563785969586687487667526481†83747271†
Surogate Rune 26B-A4B v3 (v1.6 pool)self-hosted · S+P+L1+L2+L37559474336473737454433374538195237554340502333
Piccolo 27Bself-hosted · not measured90798476888565887686798394779079628584†87908880†
Jeff-1.0-Largeself-hosted · S+P (draw v1.6-fastlane-20261009)99†718967†649156†—98†49†——————77†84†62†———59†
Lychee-27Bself-hosted · not measured—767779†7888—58†—39†———————67†70†———83†
decisio v0.8.0 on gemma-4-31B-itself-hosted · not measured———————————————————————
decider-12b v1, stock Gemma-4-12B-itself-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.8562485049424654615142464738324944615049405143
vLLM-SR d3-miniself-hosted · not measured—526257†5673—80†—36†———————30†0†———28†
Sieve-31B-v2self-hosted · not measured—716751†6791—51†—48†———————40†44†———61†
Vera Largeself-hosted · not measured—697381†5880—88†—38†———————49†45†———55†
iota-decide-4b-v0.3self-hosted · not measured—486958†6470—67†—54†———————51†13†———37†
swanOneself-hosted · S+PThe pinned model weights cannot currently be accessed (Hugging Face returned HTTP 401). Supplement measurements remain pending access recovery; the radar shows the existing S + P items.—578668†4885†26†45†91†42†—————————————
Winnow-12B Q8self-hosted · S+P+L1+L2+L37864405939523954534541294433414636575534384440
Bobcat 1.3self-hosted · not measured—697170†6387—62†—46†———————74†37†———66†
JPT-35B-A3Bself-hosted · S+P+L1+L2+L3865639483245342843473228441763934284633451720
Chon-1.0-12B-Fastself-hosted · not measured—567175†7888—49†—49†———————71†31†———44†
baselabs elephant-1self-hosted · not measured—537472†6272—54†—32†———————70†24†———55†
metask-jev-rain-31B-trainedself-hosted · S+P (draw v1.6-regular-20261012)—716968†6390—57†—46†———————73†36†———78†
deck-31Bself-hosted · S+P+L1+L2+L39178606955686560747156587150536860706159554560
Cygnetself-hosted · S+P+L1+L2+L37957385641413239634534274236214245535434394640
metask-jev-rain-12B-trainedself-hosted · S+P (draw v1.6-regular-20261012)—647481†7285—49†—45†———————59†25†———77†
DOPEST 1.0 31Bself-hosted · not measured—746075†6991—66†—51†———————79†42†———78†
Eikos-27Bself-hosted · S+P+L1+L2+L37568455440465644505838414626304139515343504249
Jev 1.13.0reference · API · S+P+L1+L2+L3766838442248353436363015431173420364323471921
Jeffan-31b-1710self-hosted · not measured—767158†6777—69†—45†———————76†58†———67†
Kev 27Bself-hosted · S+P+L1+L2+L37769475038424244644441454420265031515535482233
Jev-Omniself-hosted · S+P+L1+L2+L38656385859444537585038274928255038575547484637
GEV-26B-Decideself-hosted · S+P+L1+L2+L37852333731423742372831233028213529422140282725
decisio v0.8.0 on gemma-4-12B-itself-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.7956395539423338634129253435253744485638404227
RSI-Jev v6.1-VL 27Bself-hosted · S+P (draw v1.6-regular-20261010-a2)—5783†83†517244†90†42†32———————53————0†
Nagi-ENORMOUS 27Bself-hosted · S+P (draw v1.6-regular-20261012)—626675†7784—58†—22†———————50†23†———39†
StartLux-Decision v1.1 27B AutoThinkself-hosted · S+P (draw v1.6-regular-20261012)—667187†6184—75†—54†———————68†61†———72†
Jev 1.14.0reference · API · S+P (draw v1.6-regular-20261012)—637373†4275—51†—0†———————39†0†———44†
metask-jev-rain-31B-trained service v3self-hosted · S+P (draw v1.6-regular-20261012)—746874†6382—55†—65†———————80†43†———78†
decisio v0.9.0 on gemma-4-12B-itself-hosted · S+P+L1+L2+L38156415538403141624128293634253744485637404331
Bespoke Nimble 9B v3self-hosted · S+P+L1+L2+L3806145503544303056412774032173623324260542037
JEV-27Bself-hosted · S+P+L1+L2+L3716236422341363754393337439194737454737432648
torchcast-decision-12bself-hosted · S+P+L1+L2+L38665496447483953645044474544335440685354445844
Messier One v0.4self-hosted · S+P (draw v1.6-regular-20261012)—516862†7072—65†—39†———————32†18†———55†
shisa-de-1self-hosted · S+P (draw v1.6-regular-20261012)—456176†4588—52†—30†———————52†25†———50†
Clefself-hosted · S+P+L1+L2+L38265504927493846554533434321303136524437492336
Wald 4B v2.1self-hosted · S+P (draw v1.6-regular-20261010-a2)—5072†75†345543†54†54†49———————25————12†
Hopper 12B trainedself-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.7954505142483339544232233828163632484734322434
Blink v0.3 26B-A4Bself-hosted · S+P (draw v1.6-regular-20261010-a2)—4654†90†457054†77†39†25———————39————0†
Decision 2.0 Vega 27Bself-hosted · S+P+L1+L2+L36864394923362630353641304418222930465121432140
René-1 31B FP8self-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.7567453942464031553735375413254433474342462043
AutoJev-27B (denis-pplx, Qwen3.8-27B)self-hosted · S+P+L1+L2+L37459413727353332513830244216182736424832452527
OpenJev (thinking, BF16)self-hosted · S+P+L1+L2+L38778595431553160646653506543536046536251564142
Jebadiah 27Bself-hosted · S+P+L1+L2+L36564364730342239514129183718113133494629441726
decider chat on Gemma-4-31B-itself-hosted · S+P+L1+L2+L37564585434535045565945395245324344514443444449
NInfer Qwen3.8-Flash-Next mixedself-hosted · S+P+L1+L2+L37253415024333033553328164320273635543138432520
Metask rain 4Bself-hosted · S+P (draw v1.6-fastlane-20261009)100†486574†606261†—30†8†——————0†41†0†———73†
FrontiersMind Lumma-fev-27Bself-hosted · not measured—556469†5176—62†—26†———————33†2†———55†
Respan Span-02self-hosted · S+P (draw v1.6-regular-20261012)—605057†6180—53†—37†———————70†29†———72†
NInfer Qwen3.8-27B NVFP4self-hosted · S+P+L1+L2+L3725545513540322341423322382217183758372643178
decider-35b-a3bself-hosted · S+P+L1+L2+L377523129164523233724291644952422413624381428
Diffusion Jevself-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.8259515749653741524448294345364944635050513816
SPX-CD-Omniself-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.8053354046362932504441344426314822445635382434
KnowLine 31B-Gen1self-hosted · not measured—817268†7588—63†—46†———————80†53†———67†
Amazon Strands Decider 26B-A4Bself-hosted · not measured—436849†4969—59†—56†———————32†15†———40†
Turion Jev 31Bself-hosted · not measured—757250†6283—68†—54†———————71†61†———72†
KnowLine 4B-Gen5.5self-hosted · not measured—496446†2776—92†—15†———————31†2†———33†
JevOneself-hosted · S+P+L1+L2+L37349373522331021323122102010920243040183210
Seb-9B-v2self-hosted · not measured—446574†3168—86†—33†———————19†5†———22†
pplx-decider-flash-v1.2self-hosted · S+P (draw v1.6-regular-20261012)—476665†5474—86†—24†———————0†0†———22†
APUS-OpenJev-v1-9Bself-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.7448333125371822374180202561628293034351712
Aplomb 1self-hosted · S+P+L1+L2+L3The headline keeps the original published measurement. Its originally pinned Hugging Face revision (f8e2dc81) was later deleted upstream, so the language and category supplements (L1, L2, L3) were measured on 12 Oct 2026 with the current upstream revision d6743291 (all file hashes recorded) on an isolated pod. That revision reproduces the original answers: 98 % agreement on the 300 public items (published bar 95 %) and 97.5 % on all 1,500 headline items. Each supplement carries the original P300 answers by stable item identity; they count once. All new (sealed) supplement items were answered; the only failed requests (3 public items, GPU out-of-memory) are covered by the original P300 answers. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.694643362641212244272214242811272725433338132
JADEself-hosted · S+P+L1+L2+L320000000000000000000000
Tern-35B-A3Bself-hosted · not measured—436758†3676—56†—13†———————33†18†———21†
cribl-decision-1.0-9Bself-hosted · not measured77466540426834475848314445486562394257†60494034†
AJev lm6qself-hosted · not measured—746876†6784—64†—43†———————77†59†———65†
Saif 4Bself-hosted · not measured66445851275626586634304651474954362162†67644249†
Arcbright Elf 1 12Bself-hosted · S+P (draw v1.6-regular-20261012)—415360†5779—65†—30†———————42†0†———33†
Plumb-4Bself-hosted · S+P+L1+L2+L3724535453630382732342583240142431252827392220
cribl-decision-2self-hosted · not measured71446737405616426441274954505763403465†65643535†
Granite Switch 4.1 30Bself-hosted · not measured84516359416129545953346458266363226278†59815924†
Messier One v0.2self-hosted · not measured———————————————————————
spark-s1-4b-v6self-hosted · S+P+L1+L2+L380463331182891737352202223015292023102627
Bev / Bonsai 27Bself-hosted · S+P+L1+L2+L367462635193014172522160900024286212746
Clef-Flashself-hosted · S+P+L1+L2+L3694440318349194129231810201112232323361311
Quyet-1.0-Mediumself-hosted · S+P+L1+L2+L360443332254225153532238152002933351921351214
Open-Jev 9Bself-hosted · S+P+L1+L2+L370474048234426212635180245142819351530381417
LitJevself-hosted · S+P+L1+L2+L369493439273630233434280242210103239402930131
Bespoke Nimble 9Bself-hosted · S+P+L1+L2+L36645413635411021372619423912625263940321411
SimpleJev Qwen3.8-27B (self-hosted, v1.6 pool)self-hosted · S+P+L1+L2+L38062445234414329614133243925293241495031502926
Qwen3.5-9B Jev-like data-mix v2self-hosted · S+P+L1+L2+L37146292418376133531120253091830382424321
levself-hosted · S+P+L1+L2+L35541282022221222282229822112141728192437537
JPT-4Bself-hosted · S+P (draw v1.6-regular-20261012)—367563†2775—70†—30†———————11†14†———33†
jev-localself-hosted · S+P+L1+L2+L37146433121312122333217720127212910342529115
Aleria Judge 1.1.1self-hosted · not measured—706769†6585—58†—49†———————68†28†———53†
decider-4b v2self-hosted · S+P+L1+L2+L36242412511288222132320181318183222242911
Jeffan-4B-2010self-hosted · not measured—396676†3063—83†—14†———————20†0†———22†
ClassOne Qwen 3.5 9Bself-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.714538512636231930432301781411301718401331
Wald 4Bself-hosted · S+P (draw v1.6-fastlane-20261009)95†435152†355451†—47†0†——————8†26†14†———15†
Open-Jev 27B v1.1self-hosted · S+P+L1+L2+L3645544412240242144332837388112311423634372423
GLiClass v4 9Bself-hosted · not measured—365865†5980—71†—45†———————14†0†———39†
metask-jev-4bself-hosted · S+P+L1+L2+L37340303415211619282412915409112316122640
Kahn1 4Bself-hosted · S+P+L1+L2+L3674036462835253029452553525211433192841302020
JevK5 v0.3self-hosted · S+P+L1+L2+L36743404229192013283018028278162817223429318
deck-4B v1.0self-hosted · S+P+L1+L2+L367433745322221192930220312651323182736291917
Seb-9Bself-hosted · not measured———————————————————————
Decision-4Bself-hosted · S+P+L1+L2+L36143271903107202226261000211713152500
decisor-4bself-hosted · S+P (draw v1.6-fastlane-20261009)100†367053†445824†—45†0†——————50†12†11†———13†
janus 4Bself-hosted · S+P+L1+L2+L36340352315199152120162259001129171520107
Jebadiah 9B v2.1self-hosted · S+P (draw v1.6-regular-20261012)—356745†3270—62†—25†———————3†5†———22†
JevK5 v0.2.0self-hosted · S+P+L1+L2+L365362636212123152420601013742421191826132
Decision 2.0 Lux 9Bself-hosted · S+P (draw v1.6-regular-20261012)—316467†4067—75†—14†———————19†9†———33†
Verdin-4Bself-hosted · S+P (draw v1.6-regular-20261012)—315651†4054—69†—52†———————34†5†———21†
TypeCastLM 1.4.0self-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.57382425822161728181201790517961519133
blink-4bself-hosted · S+P (draw v1.6-regular-20261012)—317351†3179—66†—33†———————2†14†———22†
Imajev-9Bself-hosted · S+P (draw v1.6-regular-20261012)—336349†4065—64†—29†———————9†0†———15†
jiwo 4bself-hosted · S+P (draw v1.6-regular-20261010-a2)—3053†78†344341†65†20†12———————12————0†
Imajev-4Bself-hosted · S+P+L1+L2+L351373027924019121014090008101352910
Jev-Style-2Bself-hosted · not measured573336262938214228149291621112710044†4257272†
Sors-2Bself-hosted · not measured6440422339433134483815546913244330048†58481744†
Uniphore-4B-Deciderself-hosted · S+P (draw v1.6-regular-20261012)—327152†2570—74†—29†———————0†18†———15†
Malkuth-4Bself-hosted · S+P+L1+L2+L3493721173117177540700000422000
Hopperself-hosted · S+P+L1+L2+L35634312215220151612130141230121517151100
Decisio v0.9.0 on Qwen3.6-35B-A3Bself-hosted · not measured—315736†3373—54†—17†———————32†22†———36†
Clef-omniself-hosted · S+P (draw v1.6-regular-20261010-a2)—3554†74†365224†51†17†18———————45————0†
JEB-35B-A3Bself-hosted · S+P (draw v1.6-regular-20261012)—294942†4064—64†—17†———————30†7†———24†
jqvself-hosted · S+P+L1+L2+L3703916241829817151721010501872716252903
CoCo-Decision-4Bself-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.5136262321171822281716014902724162025121
Decision 4B v1.2self-hosted · S+P+L1+L2+L3653636321625162227342102432222728203534251214
Manchego v2.1self-hosted · S+P+L1+L2+L3503325110150191281702080212196473
Standard One 3B SH v4self-hosted · not measured—324763†2769—61†—24†———————26†0†———20†
Granite Switch 4.1 8Bself-hosted · not measured6234463732542240472322732293349203836†40662719†
Raw Qwen3 4B Instruct 2507 direct logitsself-hosted · S+P+L1+L2+L359341731532017244170103192142519213104
Standard One 3B SHself-hosted · S+P (draw v1.6-regular-20261012)—287270†2077—68†—22†———————20†0†———21†
Jebadiah 9B v2self-hosted · S+P (draw v1.6-regular-20261012)—276742†3167—65†—0†———————6†18†———24†
reflex 4Bself-hosted · S+P+L1+L2+L356333324413113151211061117017172941804
typecastlmself-hosted · S+P+L1+L2+L3542930196246112423130571081514820133
iota-decide-4bself-hosted · S+P (draw v1.6-regular-20261012)—247032†1668—64†—15†———————6†0†———16†
Decision 4B v1.1self-hosted · S+P+L1+L2+L3593329391224102124411701519111822727282255
Standard One 8B SHself-hosted · S+P (draw v1.6-regular-20261010-a2)—3044†77†414523†43†0†9———————4————0†
kev 8Bself-hosted · S+P+L1+L2+L351352426922087070040002092700
Compass 0.2.0self-hosted · S+P (draw v1.6-regular-20261012)—104525†1634—22†—17†———————0†0†———36†
pjev-2bself-hosted · S+P (draw v1.6-regular-20261012)—255348†3552—57†—27†———————9†23†———23†
Amber 4Bself-hosted · not measured572878†173029038†35†23†7†16331†3651290†36†52†474020†
ArseneLupin v1.1self-hosted · S+P (draw v1.6-regular-20261012)—166122†2040—44†—7†———————21†14†———5†
Standard One 8Bself-hosted · S+P+L1+L2+L362382217201712426197012304023142334610
Arbiter-4Bself-hosted · S+P (draw v1.6-regular-20261012)—275950†2066—76†—21†———————0†0†———24†
Decider-4Bself-hosted · S+P (draw v1.6-regular-20261012)—216224†1354—87†—25†———————4†4†———32†
SemIf, formerly OpenJevself-hosted · S+P+L1+L2+L3583015162225101415901900011157112000
RYOTIDE-Qwen9self-hosted · S+P (draw v1.6-fastlane-20261009)92†197045†296015†—13†0†——————0†1†15†———28†
ClassOne Gemma 4 E4Bself-hosted · S+P (draw v1.6-regular-20261010-a2)—2258†50†502125†57†31†0———————34————0†
JEV Qwen3.5-9B Base NVFP4self-hosted · S+P+L1+L2+L34027241261611211700700025130705
DecisionNode Edge 0.9.2self-hosted · not measured—204766†4869—41†—33†———————12†0†———16†
baselabs joey-1self-hosted · not measured442837296300314508180161380037†353420†
ClassOne Qwen 3.5 4Bself-hosted · S+P (draw v1.6-regular-20261010-a2)—2155†60†312337†59†27†1———————9————0†
system-oneself-hosted · S+P+L1+L2+L3513162743671719600700101618011300
Jez-large-1.0.0self-hosted · not measured—94823†2362—61†—12†———————1†0†———8†
Chon-1.0-12Bself-hosted · S+P (draw v1.6-regular-20261012)—63429†6250—24†—0†———————28†0†———44†
RYOTIDE-Qwenself-hosted · S+P (draw v1.6-regular-20261012)—227517†3373—77†—25†———————33†33†———26†
Nomaself-hosted · S+P (draw v1.6-regular-20261012)—205722†1931—79†—25†———————1†0†———44†
Raw Qwen3 8B direct logitsself-hosted · S+P+L1+L2+L3543071503001117718816000122751400
local-jev Qwen3.5-4Bself-hosted · S+P+L1+L2+L33515111105028030000078001100
OpenSourceJevself-hosted · S+P+L1+L2+L33320185211050000000000001000
Decision 2Bself-hosted · S+P+L1+L2+L320147001000000000003000000
open-alternative-jevself-hosted · S+P+L1+L2+L34324203014373911003007274401
cribl-decision-2-miniself-hosted · not measured24327211180111008000510011†183100†
Krino-2Bself-hosted · S+P (draw v1.6-regular-20261012)—161221†4345—19†—22†———————0†0†———12†
Liquid AI d1-3Bself-hosted · S+P (draw v1.6-regular-20261010-a2)—1434†55†3060†41†29†0———————1————0†
Open-Jev 2Bself-hosted · S+P+L1+L2+L339181116171009712000000003600
Imajev-2Bself-hosted · S+P (draw v1.6-regular-20261012)—54720†1742—0†—34†———————0†0†———19†
decider-2bself-hosted · S+P+L1+L2+L34225770142432802300581200700
D1A on Gemma 4 E4Bself-hosted · not measured—84142†3143—41†—2†———————0†2†———17†
Nemotron Diffusion 8Bself-hosted · S+P+L1+L2+L330160016190060004012001800300
Nev-2Bself-hosted · not measured———————————————————————
kev 4Bself-hosted · S+P+L1+L2+L344252811014051907016000013001600
DecisionNode Edge Mini 0.9.2self-hosted · not measured—33830†4546—41†—19†———————0†3†———22†
Malkuth-2Bself-hosted · S+P+L1+L2+L33514202600000000000000400
Qwen3-Reranker-4Bself-hosted · S+P+L1+L2+L340000300000000000000000
jiwo 0.8bself-hosted · S+P (draw v1.6-regular-20261010-a2)—034†59†2570†55†0†0———————0————0†
ZeroEntropy zerank-2self-hosted · S+P+L1+L2+L300000000000000000000000
Vega 4Bself-hosted · S+P (draw v1.6-regular-20261010-a2)—824†15†43422†42†10†6———————0————0†
EXAONE-4.0-1.2B-JEV v0.3self-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.56000000000000000000000
Raw Phi-4 mini direct logitsself-hosted · S+P+L1+L2+L313265000010007000010001502
SimpleJevself-hosted · S+P+L1+L2+L3681100000000000000000960
Jeetu Decider 2Bself-hosted · not measured—060†2023—42†—0†———————0†0†———18†
D1A on Gemma 4 E2Bself-hosted · not measured—0536†326—37†—0†———————0†0†———12†
smalljev semantic-v9self-hosted · S+P+L1+L2+L300000000000000000000000
Gutsy 0.8B v0.3self-hosted · S+P (draw v1.6-regular-20261010-a2)—014†27†1220†28†0†0———————0————0†
Vega-U (Qwen3.5-4B reader)self-hosted · S+P (draw v1.6-regular-20261012)—000†1010—0†—3†———————0†0†———0†
PurpleMIST-Mini-1.0self-hosted · S+P (draw v1.6-regular-20261012)—000†00—0†—1†———————0†2†———0†
kev 0.6Bself-hosted · S+P+L1+L2+L310000000000000000000000
jul fastself-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.00000000000000000000000
OpenDecisionself-hosted · S+P+L1+L2+L301000000000000000000000
ClassOne Gemma 4 E2Bself-hosted · not measuredThis model's server gives different answers after each restart, so a supplement run cannot reproduce the row's original answers (87.7 % agreement on the 300 public items, below the 95 % bar). Cause (verified 12 Oct 2026): the published model has no trained embeddings for its 12 decision delimiter tokens, so they are randomly initialised at every server start. Its language and category supplements are not published; the radar shows the existing S + P items.———————————————————————
Qwen3.5-0.8B Decision Modelself-hosted · not measured———————————————————————
Deem 0.8B v1self-hosted · S+P+L1+L2+L350000100000000000000800
Decision Fastself-hosted · S+P+L1+L2+L300000000000000000000000
Quyet-1.0-Small-ENself-hosted · S+P+L1+L2+L300000000000000000000000
ClassOne Qwen 3.5 2Bself-hosted · S+P (draw v1.6-regular-20261010-a2)—30†18†1800†39†0†0———————11————0†
Laya typed-decisionsself-hosted · S+P+L1+L2+L300000000000000000000000
Vega-U (Qwen3.5-0.8B reader)self-hosted · S+P (draw v1.6-regular-20261012)—000†00—6†—7†———————0†0†———0†
WaterSheepself-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.00000000000000000000000
openJev Verdict 1.4self-hosted · S+P+L300000000000000000000000
Raw Qwen3 0.6B direct logitsself-hosted · S+P+L1+L2+L30000000010000000130400800
lev-350mself-hosted · S+P+L1+L2+L300000000000000000000000
Raw Qwen3 1.7B direct logitsself-hosted · S+P+L1+L2+L3100060000200000000000000
kev 0.5Bself-hosted · S+P+L1+L2+L300000000000000000000000
Jez-small-1.0.0self-hosted · not measured—000†00—4†—0†———————0†0†———0†
openJev Verdictself-hosted · S+P+L1+L2+L300000000000000000000000
Quyet-1.0-Smallself-hosted · S+P+L1+L2+L300000000000000000000000
Reflex Instinct 0.6Bself-hosted · not measured———————————————————————
watt-flash-0.1self-hosted · S+P+L1+L2+L300000000000000000000000
Quyet-1.0-Tinyself-hosted · S+P+L1+L2+L300000000000000000000100
Phocinae-Largha-150M-v1self-hosted · S+P (draw v1.6-regular-20261012)—000†00—0†—0†———————0†0†———0†
Liquid AI d1-omni-600Mself-hosted · S+P (draw v1.6-regular-20261010-a2)—00†0†000†0†0†0———————0————0†
Vega 0.8Bself-hosted · S+P (draw v1.6-regular-20261010-a2)—00†22†000†5†0†0———————0————0†
open-jev-deberta-v3-largeself-hosted · S+P+L1+L2+L3———————————————————————
modernbert-jev v2self-hosted · not measured———————————————————————
Julia-1self-hosted · S+P (draw v1.6-regular-20261012)—003†022—0†—0†———————0†0†———1†
verdict-smallself-hosted · S+P+L1+L2+L300000000000000000000001
Tacet Sonataself-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.20000000000000000000000
Layaself-hosted · S+P+L1+L2+L300000000000000000000000
Mixedbread mxbai-rerank-base-v2self-hosted · S+P+L1+L2+L300000000000000000000000
Laya multilingualself-hosted · S+P+L1+L2+L3000000000000005141230000
Certo v1self-hosted · S+P+L1+L2+L300000000000000000000000
CLM-8Bself-hosted · S+P+L1+L2+L300000000000000000000000
BAAI bge-reranker-v2-m3self-hosted · S+P+L1+L2+L300000000000000000000000
Alibaba GTE Reranker ModernBERT-baseself-hosted · S+P+L1+L2+L300000000000000000000000
Mirrorself-hosted · S+P+L1+L2+L300000000000000000000000
Open Jev JSON Canvasself-hosted · S+P+L1+L2+L37665554247563755605551403446443341504754514953
APUS-OpenJev-v1-35B-A3Bself-hosted · S+P+L1+L2+L3Language and category cells use the unchanged original S+P observations and the newly completed frozen supplemental items (L1, L2, L3), answered with the original model and serving setup on an isolated pod. Each supplement carries the same original P300 answers by stable item identity; they count once in the union. Failed requests stay scored as failures and are not imputed. Headline, rank, Composite and original pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.8254374047402725414734142333123132373935371710
Bongard-miniself-hosted · S+P (draw v1.6-regular-20261012)—43413†2643—43†—0†———————0†0†———20†
jevora-9b-v1self-hosted · S+P (draw v1.6-regular-20261012)—143939†2358—65†—0†———————0†0†———16†
Kev-9B v2self-hosted · S+P (draw v1.6-regular-20261012)—304443†3375—72†—19†———————19†32†———22†
Nagi-HUGE 12Bself-hosted · S+P (draw v1.6-regular-20261012)—537370†5784—72†—12†———————42†7†———67†
Nemotron Deciffusion 8B v2self-hosted · S+P (draw v1.6-regular-20261012)—506867†5460—67†—22†———————27†7†———32†
PurpleMIST-Flash-1.0self-hosted · S+P (draw v1.6-regular-20261012)—355733†4877—57†—27†———————14†0†———16†
RYOTIDE-Gemmaself-hosted · S+P (draw v1.6-regular-20261012)—304360†4765—54†—31†———————8†6†———38†
TDE general large v0.1self-hosted · S+P (draw v1.6-regular-20261012)—000†00—0†—0†———————0†0†———0†
AutoJev-27B (RTX PRO 6000)self-hosted · S+P+L1+L2+L3Historical headline · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells come from a new measurement on the current item pools (S+P, L1, L2, L3), run with the original pinned model and serving setup on an isolated machine; each item was sent once and failed requests stay scored as failures. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.7460423827353332514030224015182837464832452527
Bosun v3.1 0.6Bself-hosted · S81+P299+L1+L2+L3Historical headline · measured language cells40000000000000000000700
GLiNER2self-hosted · S+P+L1+L2+L3Historical headline · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells come from a new measurement on the current item pools (S+P, L1, L2, L3), run with the original pinned model and serving setup on an isolated machine; each item was sent once and failed requests stay scored as failures. About 1,000 issued items were lost to out-of-memory kills and a machine shutdown; they stay counted as spent and are never re-asked, so 17 of 23 languages rest on 50 to 59 answered items instead of the 60-item target. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.110000000140000000000120010
GLiNER2 largeself-hosted · S+P+L1+L2+L3Historical headline · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells come from a new measurement on the current item pools (S+P, L1, L2, L3), run with the original pinned model and serving setup on an isolated machine; each item was sent once and failed requests stay scored as failures. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.791000000000970001008800
GLiNER2.5 multiself-hosted · S+P+L1+L2+L3Historical headline · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells come from a new measurement on the current item pools (S+P, L1, L2, L3), run with the original pinned model and serving setup on an isolated machine; each item was sent once and failed requests stay scored as failures. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.00000000000000000000400
GLiNER2.5 smallself-hosted · P+L1+L2+L3Historical headline · measured language cells00000000000000000000000
jeffself-hosted · S+P+L1+L2+L3Historical headline · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells come from a new measurement on the current item pools (S+P, L1, L2, L3), run with the original pinned model and serving setup on an isolated machine; each item was sent once and failed requests stay scored as failures. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.00000000000000000000000
Jobe Qwen3.5-4Bself-hosted · S+P+L1+L2+L3Catalogue entry · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells come from a new measurement on the current item pools (S+P, L1, L2, L3), run with the original pinned model and serving setup on an isolated machine; each item was sent once and failed requests stay scored as failures. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.58301416125491415801900011156132000
mica-v01-4bself-hosted · S+P+L1+L2+L3Catalogue entry · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells come from a new measurement on the current item pools (S+P, L1, L2, L3), run with the original pinned model and serving setup on an isolated machine; each item was sent once and failed requests stay scored as failures. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.5538322232602218264125441392230122100
Needle 3self-hosted · S+P+L1+L2+L3Historical headline · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells come from new runs of the supplements L1, L2 and L3 (all items) on the original pinned closed engine, approved by the benchmark owner, on an isolated machine, plus 281 S+P items: 153 from an earlier partial run and 128 never-asked items chosen by their use-case label ('other') to complete that radar spoke (these items also count in the other cells, so the S+P part is not a random sample); the rest of S+P was not run. The engine abstains on about 30 % of items and returns invalid UTF-8 on about 7 % (mostly non-Latin scripts); both count as unanswered. Its chance-corrected competence is below zero in every cell, which the table shows as 0. Its answers to the same public items differ between machines for about 10 % of items. 22 of 23 languages rest on 19 to 59 answered items instead of the 60-item target; the radars are complete. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.00000000000†00000†00†000†00†
Needle 3, options as toolsself-hosted · S+P+L1+L3Historical headline · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells combine an earlier partial run on the current pools (175 S+P items) with a new run of the supplements L3 (1,382 of 1,418 items) and L1 (329 of 654) on the original pinned closed engine, approved by the benchmark owner, on an isolated machine; L2 and the rest of S+P were not run before the machine's time limit. The engine abstains on about 9 % of items and returns invalid UTF-8 on about 7 % (mostly non-Latin scripts); both count as unanswered. Its chance-corrected competence is below zero in every cell, which the table shows as 0. Its answers to the same public items differ between machines for about 10 % of items. 18 of 23 languages rest on 24 to 59 answered items instead of the 60-item target; the radars are complete. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.0040000000000000†0000030†
NInfer Qwen3.8-27B NVFP4 (T=1.5)self-hosted · S+P+L1+L2+L3Historical headline · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells come from a new measurement on the current item pools (S+P, L1, L2, L3), run with the original pinned model and serving setup on an isolated machine; each item was sent once and failed requests stay scored as failures. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.715039512432292038372715371918183351262538145
OpenJev (DiffusionGemma 26B-A4B NVFP4, razorback16)self-hosted · S+P+L1+L2+L3Catalogue entry · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells come from a new measurement on the current item pools (S+P, L1, L2, L3), run with the original pinned model and serving setup on an isolated machine; each item was sent once and failed requests stay scored as failures. This model samples stochastically: on the 300 public items its answers in the supplement passes agree with its S+P pass on 86-89 %; those public items count once, from the S+P pass. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.6054464149462937424437303023273128332735313334
reflex-27bself-hosted · S+P+L1+L2+L3Historical headline · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells come from a new measurement on the current item pools (S+P, L1, L2, L3), run with the original pinned model and serving setup on an isolated machine; each item was sent once and failed requests stay scored as failures. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.685634382736322047492920282311183137374030135
Surogate Rune 26B-A4B v3 (RTX PRO 6000)self-hosted · S+P+L1+L2+L3Historical headline · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells combine its S+P run on the current pools (6 Oct 2026) with new L1, L2 and L3 runs on the same pinned model and serving setup on an isolated machine; each item was sent once and failed requests stay scored as failures. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.7559474336473737454433374538195237554340502333
Vonself-hosted · S+P+L1+L2+L3Historical headline · measured language cellsThis historical row keeps its published v1.5 headline. Its language and category cells come from a new measurement on the current item pools (S+P, L1, L2, L3), run with the original pinned model and serving setup on an isolated machine; each item was sent once and failed requests stay scored as failures. Headline, rank, Composite and pricing are unchanged. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3.00000000000000000000000
metask-jev-rain-12Bself-hosted · S+P (draw v1.6-fastlane-20261009)99†647978†558367†—71†42†——————67†65†47†———71†

Intelligence gate and Noul decisiveness

The Composite applies a soft gate below Intelligence 50 (× (I/50)²). Rows under it show their value with a "below gate" label. The decisive rate is the share of valid Noul answers with P(yes) ≥ 0.80 or ≤ 0.20 (for systems that return only a label, every valid label counts as decisive); only decisive answers count as answers. Accuracy among decisive answers is a diagnostic and does not enter any score.

SystemIntelligenceChoiceNoulScoreNoul decisive rateAccuracy among decisive
Torchcast Decision 4B67.1nativenativenative100 %86 %
H2O-Lightning-4B v1.160.0nativenativenative100 %85 %
Wald 26B-A4B65.5nativenativenative100 %82 %
Torchcast Decision 4B60.3nativenativenative100 %84 %
Fastino GLiDE 1.276.4nativenativenative86 %98 %
Bobcat Flash 1.265.6nativenativenative89 %92 %
H2O-Lightning-31B v1.0.779.3nativenativenative100 %92 %
StartLux-Decision v1.1 27B Instant73.2nativenativenative82 %99 %
baselabs mammoth-176.2nativenativenative89 %97 %
Decisio 12B tuned commit62.8nativenativenative100 %87 %
Jeff-1.1-Medium74.5nativenativenative100 %89 %
Philly Decider v256.1nativenativenative100 %80 %
Perplexity Decider v1.1 27B75.6nativenativenative91 %95 %
Chon-1.0-E4B55.5nativenativenative100 %80 %
vLLM-SR d3-turbo74.8nativenativenative87 %96 %
vLLM-SR d372.9nativenativenative86 %96 %
KnowLine 26B-A4B-Gen2.565.2nativenativenativeunavailableunavailable
Torchcast Decision 4B54.2nativenativenative100 %78 %
KnowLine 35B-A3B-Gen1.560.3nativenativenative75 %95 %
basal-omni-1.0-ultra74.1nativenativenative92 %94 %
Xor 26B-A4B57.5nativenativenative81 %94 %
torchcast-decision-27b70.4nativenativenative83 %97 %
Arcbright Elf 1 31B71.9nativenativenative100 %91 %
pplx-decider-v1.271.7nativenativenative84 %96 %
Quyet-1.0-Large73.4nativenativenative91 %93 %
Nimble-v4-max71.0nativenativenative84 %97 %
decider-12b v261.9nativenativenative99 %83 %
Xevn-4B52.1nativenativenative100 %79 %
Surogate Rune 26B-A4B v356.1nativenativenative73 %97 %
Jeff-1.0-Large71.9nativenativenativenot supported—
Lychee-27B74.3nativenativenative92 %94 %
decisio v0.8.0 on gemma-4-31B-it70.4nativenativenative94 %96 %
decider-12b v1, stock Gemma-4-12B-it59.6nativenativenative96 %87 %
vLLM-SR d3-mini50.8nativenativenative77 %91 %
Sieve-31B-v268.7nativenativenative82 %98 %
Vera Large67.8nativenativenative82 %95 %
iota-decide-4b-v0.352.5nativenativenative100 %79 %
swanOne56.2nativenativenative81 %93 %
Winnow-12B Q859.5nativenativenative86 %92 %
Bobcat 1.368.1nativenativenative92 %94 %
JPT-35B-A3B51.4nativenativenative74 %94 %
Chon-1.0-12B-Fast60.3nativenativenative100 %78 %
baselabs elephant-155.8nativenativenative78 %91 %
metask-jev-rain-31B-trained70.7nativenativenative100 %88 %
deck-31B73.0nativenativenative96 %94 %
Cygnet54.8nativenativenative81 %92 %
metask-jev-rain-12B-trained66.5nativenativenative100 %86 %
DOPEST 1.0 31B73.9nativenativenative100 %90 %
Eikos-27B64.4nativenativenative81 %98 %
Jeffan-31b-171071.6nativenativenative87 %94 %
Kev 27B64.2nativenativenative80 %97 %
Jev-Omni55.5nativenativenative74 %94 %
GEV-26B-Decide49.8below gatenativenativenative70 %96 %
decisio v0.8.0 on gemma-4-12B-it52.3nativenativenative82 %91 %
RSI-Jev v6.1-VL 27B61.6nativenativenative83 %91 %
Nagi-ENORMOUS 27B62.4nativenativenative84 %91 %
StartLux-Decision v1.1 27B AutoThink66.7nativenativenative77 %97 %
metask-jev-rain-31B-trained service v372.6nativenativenative100 %89 %
decisio v0.9.0 on gemma-4-12B-it52.3nativenativenative81 %91 %
Bespoke Nimble 9B v356.5nativenativenative87 %89 %
JEV-27B56.8nativenativenative79 %95 %
torchcast-decision-12b60.5nativenativenative96 %86 %
Messier One v0.452.0nativenativenative100 %78 %
shisa-de-149.9below gatenativenativenative76 %90 %
Clef59.1nativenativenative82 %94 %
Wald 4B v2.152.0nativenativenative100 %78 %
Hopper 12B trained51.7nativenativenative79 %93 %
Blink v0.3 26B-A4B49.2below gatenativenativenative67 %95 %
Decision 2.0 Vega 27B60.4nativenativenative73 %99 %
René-1 31B FP861.7nativenativenative78 %98 %
AutoJev-27B55.6nativenativenative70 %99 %
OpenJev70.2nativenativenative96 %93 %
Jebadiah 27B57.5nativenativenative74 %98 %
decider chat on Gemma-4-31B-it58.5nativenativenative80 %98 %
NInfer Qwen3.8-Flash-Next mixed49.0below gatenativenativenative75 %93 %
Metask rain 4B51.1nativenativenativenot supported—
FrontiersMind Lumma-fev-27B55.5nativenativenative73 %93 %
Respan Span-0260.9nativenativenative90 %87 %
NInfer Qwen3.8-27B NVFP451.8nativenativenative82 %89 %
decider-35b-a3b48.7below gatenativenativenative74 %91 %
Diffusion Jev54.9nativenativenative93 %85 %
SPX-CD-Omni48.8below gatenativenativenative68 %93 %
KnowLine 31B-Gen177.4nativenativenative91 %97 %
Amazon Strands Decider 26B-A4B46.7below gatenativenativenative72 %91 %
Turion Jev 31B70.9nativenativenative87 %95 %
KnowLine 4B-Gen5.546.4below gatenativenativenative70 %89 %
JevOne46.0below gatenativenativenative62 %96 %
Seb-9B-v246.1below gatenativenativenative70 %91 %
pplx-decider-flash-v1.245.5below gatenativenativenative79 %87 %
APUS-OpenJev-v1-9B47.2below gatenativenativenative88 %85 %
Aplomb 145.2below gatenativenativenative68 %93 %
JADE47.9below gatenativenativenative94 %91 %
Tern-35B-A3B44.1below gatenativenativenative63 %96 %
AJev lm6q71.5nativenativenative87 %98 %
Arcbright Elf 1 12B43.8below gatenativenativenative63 %97 %
Plumb-4B42.8below gatenativenativenative71 %91 %
Messier One v0.242.4below gatenativenativenative85 %85 %
spark-s1-4b-v643.2below gatenativenativenative84 %83 %
Bev / Bonsai 27B43.8below gatenativenativenative70 %91 %
Clef-Flash42.2below gatenativenativenative68 %93 %
Quyet-1.0-Medium41.7below gatenativenativenative76 %84 %
Open-Jev 9B43.9below gatenativenativenative81 %86 %
LitJev43.3below gatenativenativenative66 %98 %
Bespoke Nimble 9B43.1below gatenativenativenative77 %87 %
SimpleJev Qwen3.8-27B57.9nativenativenative78 %96 %
Qwen3.5-9B Jev-like data-mix v241.9below gatenativenativenative71 %91 %
lev41.1below gatenativenativenative84 %83 %
JPT-4B40.8below gatenativenativenative66 %90 %
jev-local41.9below gatenativenativenative85 %82 %
Aleria Judge 1.1.168.0nativenativenative94 %93 %
decider-4b v240.1below gatenativenativenative64 %92 %
Jeffan-4B-201040.1below gatenativenativenative63 %91 %
ClassOne Qwen 3.5 9B42.2below gatenativenativenative97 %79 %
Wald 4B41.0below gatenativenativenativenot supported—
Open-Jev 27B v1.152.5nativenativenative67 %94 %
GLiClass v4 9B39.6below gatenativenativenative70 %89 %
metask-jev-4b39.2below gatenativenativenative75 %88 %
Kahn1 4B39.7below gatenativenativenative70 %87 %
JevK5 v0.338.5below gatenativenativenative66 %95 %
deck-4B v1.038.6below gatenativenativenative68 %93 %
Seb-9B38.3below gatenativenativenative57 %94 %
Decision-4B37.9below gatenativenativenative72 %88 %
decisor-4b39.6below gatenativenativenativenot supported—
janus 4B37.6below gatenativenativenative63 %94 %
Jebadiah 9B v2.137.2below gatenativenativenative60 %93 %
JevK5 v0.2.036.3below gatenativenativenative62 %93 %
Decision 2.0 Lux 9B35.8below gatenativenativenative53 %94 %
Verdin-4B35.4below gatenativenativenative65 %86 %
TypeCastLM 1.4.035.4below gatenativenativenative77 %81 %
blink-4b35.0below gatenativenativenative60 %90 %
Imajev-9B36.7below gatenativenativenative54 %93 %
jiwo 4b34.8below gatenativenativenative63 %90 %
Imajev-4B34.6below gatenativenativenative52 %96 %
Uniphore-4B-Decider34.2below gatenativenativenative72 %82 %
Malkuth-4B33.8below gatenativenativenative65 %92 %
Hopper33.6below gatenativenativenative57 %94 %
Decisio v0.9.0 on Qwen3.6-35B-A3B34.0below gatenativenativenative54 %95 %
Clef-omni34.5below gatenativenativenative65 %86 %
JEB-35B-A3B33.2below gatenativenativenative54 %94 %
jqv33.9below gatenativenativenative66 %94 %
CoCo-Decision-4B32.7below gatenativenativenative56 %96 %
Decision 4B v1.232.4below gatenativenativenative59 %94 %
Manchego v2.132.5below gatenativenativenative62 %90 %
Standard One 3B SH v434.4below gatenativenativenative67 %84 %
Raw Qwen3 4B Instruct 2507 direct logits34.0below gatenativenativenative92 %73 %
Standard One 3B SH32.8below gatenativenativenative63 %84 %
Jebadiah 9B v230.6below gatenativenativenative52 %93 %
reflex 4B30.8below gatenativenativenative58 %92 %
typecastlm29.8below gatenativenativenative57 %90 %
iota-decide-4b29.4below gatenativenativenative60 %84 %
Decision 4B v1.129.1below gatenativenativenative57 %94 %
Standard One 8B SH34.5below gatenativenativenative61 %88 %
kev 8B29.3below gatenativenativenative77 %81 %
Compass 0.2.028.9below gatenativenativenative25 %97 %
pjev-2b28.8below gatenativenativenative59 %88 %
ArseneLupin v1.128.0below gatenativenativenative42 %94 %
Standard One 8B32.8below gatenativenativenative68 %90 %
Arbiter-4B27.4below gatenativenativenative65 %84 %
Decider-4B27.1below gatenativenativenative54 %90 %
SemIf, formerly OpenJev27.1below gatenativenativenative62 %90 %
RYOTIDE-Qwen927.6below gatenativenativenativenot supported—
ClassOne Gemma 4 E4B28.2below gatenativenativenative93 %74 %
JEV Qwen3.5-9B Base NVFP429.3below gatenativenativenative53 %89 %
DecisionNode Edge 0.9.226.8below gatenativenativenative62 %83 %
ClassOne Qwen 3.5 4B27.3below gatenativenativenative86 %71 %
system-one29.3below gatenativenativenative92 %73 %
Jez-large-1.0.025.6below gatenativenativenative34 %98 %
Chon-1.0-12B27.3below gatenativenativenative100 %90 %
RYOTIDE-Qwen25.2below gatenativenativenative57 %89 %
Noma25.1below gatenativenativenative51 %88 %
Raw Qwen3 8B direct logits26.5below gatenativenativenative92 %71 %
local-jev Qwen3.5-4B24.1below gatenativenativenative35 %99 %
OpenSourceJev23.2below gatenativenativenative49 %95 %
Decision 2B22.9below gatenativenativenative39 %94 %
open-alternative-jev22.3below gatenativenativenative60 %87 %
Krino-2B22.0below gatenativenativenative83 %71 %
Liquid AI d1-3B22.1below gatenativenativenative49 %83 %
Open-Jev 2B21.8below gatenativenativenative67 %75 %
Imajev-2B21.2below gatenativenativenative38 %84 %
decider-2b20.7below gatenativenativenative68 %78 %
D1A on Gemma 4 E4B19.9below gatenativenativenative46 %84 %
Nemotron Diffusion 8B20.4below gatenativenativenative61 %77 %
kev 4B19.5below gatenativenativenative67 %82 %
DecisionNode Edge Mini 0.9.217.6below gatenativenativenative47 %79 %
Malkuth-2B15.9below gatenativenativenative53 %84 %
Qwen3-Reranker-4B16.3below gatenativenativenative25 %75 %
jiwo 0.8b15.9below gatenativenativenative36 %76 %
ZeroEntropy zerank-215.6below gatenativenativenative14 %87 %
Vega 4B15.7below gatenativenativenative69 %70 %
EXAONE-4.0-1.2B-JEV v0.314.8below gatenativenativenative40 %85 %
Raw Phi-4 mini direct logits13.7below gatenativenativenative50 %61 %
SimpleJev12.7below gatenativenativenative100 %58 %
Jeetu Decider 2B10.2below gatenativenativenative38 %76 %
D1A on Gemma 4 E2B8.9below gatenativenativenative28 %78 %
smalljev semantic-v98.7below gatenativenativenative23 %73 %
Gutsy 0.8B v0.38.6below gatenativenativenative29 %74 %
Vega-U8.5below gatenativenativenative33 %76 %
PurpleMIST-Mini-1.08.4below gatenativenativenative9 %88 %
kev 0.6B7.6below gatenativenativenative43 %69 %
jul fast7.5below gatenativenativenative30 %89 %
OpenDecision7.4below gatenativenativenative51 %68 %
ClassOne Gemma 4 E2B7.8below gatenativenativenative83 %59 %
Deem 0.8B v16.5below gatenativenativenative77 %59 %
Decision Fast6.1below gatenativenativenative14 %90 %
Quyet-1.0-Small-EN6.1below gatenativenativenative28 %64 %
ClassOne Qwen 3.5 2B5.8below gatenativenativenative87 %62 %
Laya typed-decisions5.5below gatenativenativenative1 %100 %
Vega-U5.3below gatenativenativenative20 %53 %
WaterSheep5.3below gatenativenativenative37 %69 %
openJev Verdict 1.45.0below gatenativenativenative0 %—
Raw Qwen3 0.6B direct logits5.6below gatenativenativenative100 %42 %
lev-350m4.8below gatenativenativenative8 %69 %
Raw Qwen3 1.7B direct logits5.1below gatenativenativenative100 %42 %
kev 0.5B4.6below gatenativenativenative26 %56 %
Jez-small-1.0.04.6below gatenativenativenative11 %85 %
openJev Verdict4.1below gatenativenativenative33 %60 %
Quyet-1.0-Small4.0below gatenativenativenative29 %71 %
watt-flash-0.13.8below gatenativenativenative1 %60 %
Quyet-1.0-Tiny3.7below gatenativenativenative20 %64 %
Phocinae-Largha-150M-v13.6below gatenativenativenative3 %55 %
Liquid AI d1-omni-600M3.3below gatenativenativenative27 %56 %
Vega 0.8B3.1below gatenativenativenative32 %63 %
open-jev-deberta-v3-large2.8below gatenativenativenative15 %58 %
Julia-12.4below gatenativenativenative73 %46 %
verdict-small2.3below gatenativenativenative45 %62 %
Tacet Sonata2.2below gatenativenativenative32 %56 %
Laya1.8below gatenativenativenative19 %61 %
Mixedbread mxbai-rerank-base-v21.1below gatenativenativenative0 %—
Laya multilingual0.3below gatenativenativenative72 %50 %
Certo v10.2below gatenativenativenative1 %25 %
CLM-8B0.1below gatenativenativenative51 %69 %
BAAI bge-reranker-v2-m30.1below gatenativenativenative0 %—
Alibaba GTE Reranker ModernBERT-base0.0below gatenativenativenative0 %—
Mirror0.0below gatenativenativenative92 %60 %
Open Jev JSON Canvas61.2labellabellabel100 %85 %
Jev 1.13.0 · reference63.6nativenativenative79 %98 %
Jev 1.14.0 · reference59.9nativenativenative76 %95 %
All data (296 systems)

What-If: GPU cost calculator

What-If only — the official Cost axis and all ranks use the method's reference prices.

Hosting mode

One stream uses measured median single-stream latency. More than one assumes linear batching gains that were not measured. Hardware overrides change prices only: latency is retained, performance and VRAM fit on other hardware are unverified. Unknown hardware uses H100 assumptions.

Estimated costs sorted lowest first; these are not official ranks.
SystemGPUCost per 1,000 decisionsOfficial cost (reference pricing)Ratio to official
BAAI bge-reranker-v2-m3RTX A6000$0.002701$0.022690.119×
Alibaba GTE Reranker ModernBERT-baseRTX A6000$0.003637$0.010580.344×
Raw Qwen3 0.6B direct logitsRTX A6000$0.004332$0.0033791.28×
Mixedbread mxbai-rerank-base-v2RTX A6000$0.004726$0.0210.225×
Raw Qwen3 1.7B direct logitsRTX A6000$0.004773$0.0033791.41×
lev-350mRTX 6000 Ada$0.005247$0.0046281.13×
Certo v1RTX A6000$0.006223$0.0012674.91×
JEV Qwen3.5-9B Base NVFP4RTX 5090$0.006339$0.055840.114×
Raw Phi-4 mini direct logitsRTX A6000$0.006419$0.036250.177×
kev 0.5BRTX A6000$0.006485$0.0033641.93×
verdict-smallCPU (8 vCPU)$0.007243$0.00086628.36×
Raw Qwen3 4B Instruct 2507 direct logitsRTX A6000$0.007329$0.0054521.34×
decider-4b v2RTX 5090$0.007773$0.0049791.56×
Messier One v0.2RTX 5090$0.008102$0.0048491.67×
kev 0.6BRTX A6000$0.008331$0.0033622.48×

Missing or invalid inputs show —. All costs are per useful hour of dedicated hardware; idle time is paid. Own = purchase / (years × 8,760 × utilisation) + watts / 1,000 × USD/kWh × PUE / utilisation. Rental = USD/hour / utilisation. Decisions/hour = 3,600 / median seconds × streams. Cost/1,000 = useful hourly cost / decisions/hour × 1,000. Ratio = What-If / official cost; a zero official cost has no ratio.

Own hardware assumes constant power, including idle time; GPU prices exclude host, tax, storage and maintenance. CPU defaults describe an 8-vCPU allocation. Reserved prices are assumptions, not offers. Defaults: three years, USD 0.15/kWh, PUE 1.3, utilisation 1, one stream. Edits stay in this page only.

Preset sources and assumptions

RunPod public pricing (read 2026-10-05; reserved quotes require sales).

  • H100 PCIe: RunPod Pods pricing https://www.runpod.io/pricing, read 2026-10-05: on-demand and VRAM. Purchase, watts and reserved rate are round assumptions, not vendor quotes; GPU-only purchase excludes host, tax and maintenance.
  • RTX 6000 Ada: RunPod Pods pricing https://www.runpod.io/pricing, read 2026-10-05: on-demand and VRAM. Purchase, watts and reserved rate are round assumptions, not vendor quotes; GPU-only purchase excludes host, tax and maintenance.
  • RTX PRO 6000 Blackwell: RunPod Pods pricing https://www.runpod.io/pricing, read 2026-10-05: on-demand and VRAM. Purchase, watts and reserved rate are round assumptions, not vendor quotes; GPU-only purchase excludes host, tax and maintenance.
  • RTX 5090: RunPod Pods pricing https://www.runpod.io/pricing, read 2026-10-05: on-demand and VRAM. Purchase, watts and reserved rate are round assumptions, not vendor quotes; GPU-only purchase excludes host, tax and maintenance.
  • RTX A6000: RunPod Pods pricing https://www.runpod.io/pricing, read 2026-10-05: on-demand and VRAM. Purchase, watts and reserved rate are round assumptions, not vendor quotes; GPU-only purchase excludes host, tax and maintenance.
  • CPU (8 vCPU): 8-vCPU server allocation: all prices and watts are round assumptions, set 2026-10-05; CPU — use the on-demand field for your provider quote.

What a decision costs

Every price here is US dollars per 1,000 decisions, not per 1,000 tokens. One decision is a whole typed request (state, rubric and options), not a single token.

How costs are estimated

Rows with a public tariff (per token or per request) are priced at that tariff times the tokens they used; 1 row on this board carries one. Open-weights rows have no tariff of their own: they carry the market reference price of their base model (the list price of the same weights at a large inference provider) times their tokens, marked “est.”. Most rows keep their documented v1.5 cost per 1,000 decisions; each row's basis below and in its cost cell says how it was priced. We do not use per-minute GPU rental or our own CPU time; the What-If calculator above shows that view separately.

Price rules. Only public, bookable list prices that have been in effect for at least 30 days count; a manufacturer's standard, non-promotional launch list price counts from day one, and promotions, subsidies, credits and free tiers never do. A self-hosted open-weights row is never priced below its base model's market reference price; API offerings are ranked at their own list price. A system without any eligible price is listed as unpriced. A later price change triggers a re-score with a visible note on the row.

  • Torchcast Decision 4B: ~$0.0049 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • H2O-Lightning-4B v1.1: ~$0.0052 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Wald 26B-A4B: ~$0.0093 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Torchcast Decision 4B: ~$0.0049 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Fastino GLiDE 1.2: ~$0.0249 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Bobcat Flash 1.2: ~$0.0092 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • H2O-Lightning-31B v1.0.7: ~$0.0292 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • StartLux-Decision v1.1 27B Instant: ~$0.0254 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • baselabs mammoth-1: ~$0.0268 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Decisio 12B tuned commit: ~$0.0126 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Jeff-1.1-Medium: ~$0.0134 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Philly Decider v2: ~$0.0050 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Perplexity Decider v1.1 27B: ~$0.0264 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Chon-1.0-E4B: ~$0.0035 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • vLLM-SR d3-turbo: ~$0.0266 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • vLLM-SR d3: ~$0.0266 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • KnowLine 26B-A4B-Gen2.5: ~$0.0180 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Torchcast Decision 4B: ~$0.0049 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • KnowLine 35B-A3B-Gen1.5: ~$0.0164 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • basal-omni-1.0-ultra: ~$0.0295 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Xor 26B-A4B: ~$0.0085 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • torchcast-decision-27b: ~$0.0257 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Arcbright Elf 1 31B: ~$0.0314 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • pplx-decider-v1.2: ~$0.0263 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Quyet-1.0-Large: ~$0.0296 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Nimble-v4-max: ~$0.0261 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • decider-12b v2: ~$0.0115 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Xevn-4B: ~$0.0045 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Surogate Rune 26B-A4B v3: ~$0.0092 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Jeff-1.0-Large: ~$0.0328 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Lychee-27B: ~$0.0264 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • decisio v0.8.0 on gemma-4-31B-it: ~$0.0316 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • decider-12b v1, stock Gemma-4-12B-it: ~$0.0115 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • vLLM-SR d3-mini: ~$0.0051 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Sieve-31B-v2: ~$0.0288 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Vera Large: ~$0.0294 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • iota-decide-4b-v0.3: ~$0.0051 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • swanOne: ~$0.0090 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Winnow-12B Q8: ~$0.0126 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Bobcat 1.3: ~$0.0312 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • JPT-35B-A3B: ~$0.0083 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Chon-1.0-12B-Fast: ~$0.0107 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • baselabs elephant-1: ~$0.0116 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • metask-jev-rain-31B-trained: ~$0.0290 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • deck-31B: ~$0.0316 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Cygnet: ~$0.0128 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • metask-jev-rain-12B-trained: ~$0.0145 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • DOPEST 1.0 31B: ~$0.0379 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Eikos-27B: ~$0.0292 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Jeffan-31b-1710: ~$0.0417 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Kev 27B: ~$0.0289 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Jev-Omni: ~$0.0131 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • GEV-26B-Decide: ~$0.0082 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • decisio v0.8.0 on gemma-4-12B-it: ~$0.0119 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • RSI-Jev v6.1-VL 27B: ~$0.0238 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Nagi-ENORMOUS 27B: ~$0.0264 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • StartLux-Decision v1.1 27B AutoThink: ~$0.0280 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Jev 1.14.0: ~$0.0321 est. per 1,000 decisions. TypeSafe AI public list price, USD 0.042/M input, output free: the "Current models" table of TypeSafe's models page (docs.typesafe.ai/models, fetched 16 Oct 2026 22:37:44Z, saved copy sha256 936da179...) lists jev-1.1…
  • metask-jev-rain-31B-trained service v3: ~$0.0307 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • decisio v0.9.0 on gemma-4-12B-it: ~$0.0119 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Bespoke Nimble 9B v3: ~$0.0138 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • JEV-27B: ~$0.0241 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • torchcast-decision-12b: ~$0.0285 est. per 1,000 decisions. v1.5.8 fast-lane delivery row (combined v1.5.8 preview, not yet published) cost per 1,000 decisions carried (pricing rules unchanged.
  • Messier One v0.4: ~$0.0065 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • shisa-de-1: ~$0.0092 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Clef: ~$0.0305 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Wald 4B v2.1: ~$0.0202 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Hopper 12B trained: ~$0.0130 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Blink v0.3 26B-A4B: ~$0.0092 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Decision 2.0 Vega 27B: ~$0.0271 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • René-1 31B FP8: ~$0.0421 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • AutoJev-27B: ~$0.0277 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • OpenJev: ~$0.0320 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Jebadiah 27B: ~$0.0262 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • decider chat on Gemma-4-31B-it: ~$0.0290 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • NInfer Qwen3.8-Flash-Next mixed: ~$0.0087 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Metask rain 4B: ~$0.0104 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • FrontiersMind Lumma-fev-27B: ~$0.0321 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Respan Span-02: ~$0.0354 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • NInfer Qwen3.8-27B NVFP4: ~$0.0282 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • decider-35b-a3b: ~$0.0078 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Diffusion Jev: ~$0.0091 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • SPX-CD-Omni: ~$0.0111 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • KnowLine 31B-Gen1: ~$0.0608 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Amazon Strands Decider 26B-A4B: ~$0.0086 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Turion Jev 31B: ~$0.0603 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • KnowLine 4B-Gen5.5: ~$0.0101 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • JevOne: ~$0.0081 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Seb-9B-v2: ~$0.0112 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • pplx-decider-flash-v1.2: ~$0.0050 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • APUS-OpenJev-v1-9B: ~$0.0113 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Aplomb 1: ~$0.0083 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • JADE: ~$0.0198 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Tern-35B-A3B: ~$0.0090 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • AJev lm6q: ~$0.0726 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Arcbright Elf 1 12B: ~$0.0118 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Plumb-4B: ~$0.0056 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Messier One v0.2: ~$0.0048 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • spark-s1-4b-v6: ~$0.0068 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Bev / Bonsai 27B: ~$0.0299 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Clef-Flash: ~$0.0131 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Quyet-1.0-Medium: ~$0.0054 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Open-Jev 9B: ~$0.0377 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • LitJev: ~$0.0299 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Bespoke Nimble 9B: ~$0.0284 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • SimpleJev Qwen3.8-27B: ~$0.0827 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Qwen3.5-9B Jev-like data-mix v2: ~$0.0143 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • lev: ~$0.0061 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • JPT-4B: ~$0.0051 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • jev-local: ~$0.0052 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Aleria Judge 1.1.1: ~$0.0916 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • decider-4b v2: ~$0.0050 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Jeffan-4B-2010: ~$0.0051 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • ClassOne Qwen 3.5 9B: ~$0.0108 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Wald 4B: ~$0.0096 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Open-Jev 27B v1.1: ~$0.0876 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • GLiClass v4 9B: ~$0.0093 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • metask-jev-4b: ~$0.0063 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Kahn1 4B: ~$0.0151 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • JevK5 v0.3: ~$0.0056 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • deck-4B v1.0: ~$0.0054 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Seb-9B: ~$0.0112 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Decision-4B: ~$0.0045 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • decisor-4b: ~$0.0051 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • janus 4B: ~$0.0050 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Jebadiah 9B v2.1: ~$0.0112 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • JevK5 v0.2.0: ~$0.0056 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Decision 2.0 Lux 9B: ~$0.0116 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Verdin-4B: ~$0.0060 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • TypeCastLM 1.4.0: ~$0.0049 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • blink-4b: ~$0.0054 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Imajev-9B: ~$0.0448 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • jiwo 4b: ~$0.0052 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Imajev-4B: ~$0.0055 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Uniphore-4B-Decider: ~$0.0050 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Malkuth-4B: ~$0.0056 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Hopper: ~$0.0059 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Decisio v0.9.0 on Qwen3.6-35B-A3B: ~$0.0189 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Clef-omni: ~$0.0292 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • JEB-35B-A3B: ~$0.0084 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • jqv: ~$0.0323 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • CoCo-Decision-4B: ~$0.0047 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Decision 4B v1.2: ~$0.0056 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Manchego v2.1: ~$0.0051 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Standard One 3B SH v4: ~$0.0515 est. per 1,000 decisions. ESTIMATE (*est.), base-model market reference, the same estimate as the live row standard_one_3b_sh: the model card names base_model StandardThinking/StandardOne-3B, built on mistralai/Ministral-3-3B-Instruct-2512.
  • Raw Qwen3 4B Instruct 2507 direct logits: ~$0.0055 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Standard One 3B SH: ~$0.0515 est. per 1,000 decisions. v1.5-M2 base-model reference (read 2026-10-12T00:14Z, source/public-docs/OPENROUTER-MINISTRAL-3-PRICES.json): the model card names base_model StandardThinking/StandardOne-3B, built on mistralai/Ministral-3-3B-Instruct…
  • Jebadiah 9B v2: ~$0.0112 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • reflex 4B: ~$0.0055 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • typecastlm: ~$0.0052 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • iota-decide-4b: ~$0.0056 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Decision 4B v1.1: ~$0.0056 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Standard One 8B SH: ~$0.0782 est. per 1,000 decisions. v1.5-M2 base-model reference: model card base_model StandardThinking/StandardOne-8B, built on mistralai/Ministral-3-8B-Instruct-2512 (not listed on OpenRouter as Standard One).
  • kev 8B: ~$0.0109 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Compass 0.2.0: ~$0.0085 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • pjev-2b: ~$0.0164 est. per 1,000 decisions. ESTIMATE (*est.), size-class reference, carried over unchanged from PREREG-R36.md section 3.2.
  • ArseneLupin v1.1: ~$0.0067 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Standard One 8B: ~$0.0777 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Arbiter-4B: ~$0.0044 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Decider-4B: ~$0.0044 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • SemIf, formerly OpenJev: ~$0.0056 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • RYOTIDE-Qwen9: ~$0.0135 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • ClassOne Gemma 4 E4B: ~$0.0040 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • JEV Qwen3.5-9B Base NVFP4: ~$0.0558 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • DecisionNode Edge 0.9.2: ~$0.0028 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • ClassOne Qwen 3.5 4B: ~$0.0048 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • system-one: ~$0.0114 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Jez-large-1.0.0: ~$0.0036 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Chon-1.0-12B: ~$0.0147 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • RYOTIDE-Qwen: ~$0.0060 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Noma: ~$0.0061 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Raw Qwen3 8B direct logits: ~$0.0110 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • local-jev Qwen3.5-4B: ~$0.0056 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • OpenSourceJev: ~$0.0035 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Decision 2B: ~$0.0132 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • open-alternative-jev: ~$0.0168 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Krino-2B: ~$0.0034 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Liquid AI d1-3B: ~$0.0445 est. per 1,000 decisions. ESTIMATE (*est.), size-class reference; self-hosted open weights, no bookable listing for this base.
  • Open-Jev 2B: ~$0.0103 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Imajev-2B: ~$0.0123 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • decider-2b: ~$0.0030 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • D1A on Gemma 4 E4B: ~$0.0042 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Nemotron Diffusion 8B: ~$0.0305 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • kev 4B: ~$0.0055 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • DecisionNode Edge Mini 0.9.2: ~$0.0028 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Malkuth-2B: ~$0.0035 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Qwen3-Reranker-4B: ~$0.0206 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • jiwo 0.8b: ~$0.0032 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • ZeroEntropy zerank-2: ~$0.0206 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Vega 4B: ~$0.0189 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • EXAONE-4.0-1.2B-JEV v0.3: ~$0.0253 est. per 1,000 decisions. ESTIMATE (size-class reference; self-hosted open weights, no bookable listing for this base).
  • Raw Phi-4 mini direct logits: ~$0.0363 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • SimpleJev: ~$0.0059 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Jeetu Decider 2B: ~$0.0030 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • D1A on Gemma 4 E2B: ~$0.0034 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • smalljev semantic-v9: ~$0.0203 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Gutsy 0.8B v0.3: ~$0.0032 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Vega-U: ~$0.0048 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • PurpleMIST-Mini-1.0: ~$0.0029 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • kev 0.6B: ~$0.0034 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • jul fast: ~$0.0108 est. per 1,000 decisions. ESTIMATE (size-class reference; self-hosted open weights, no bookable listing for this base).
  • OpenDecision: ~$0.0050 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • ClassOne Gemma 4 E2B: ~$0.0030 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Deem 0.8B v1: ~$0.0027 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Decision Fast: ~$0.0028 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Quyet-1.0-Small-EN: ~$0.0023 est. per 1,000 decisions. v1.5.8 fast-lane delivery row (combined v1.5.8 preview, not yet published) cost per 1,000 decisions carried (pricing rules unchanged.
  • ClassOne Qwen 3.5 2B: ~$0.0030 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Laya typed-decisions: ~$0.0038 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Vega-U: ~$0.0030 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • WaterSheep: ~$0.0111 est. per 1,000 decisions. ESTIMATE, labelled, and RUN 48'S RATE FOR THIS EXACT ROW AND BASE, CARRIED UNCHANGED: answerdotai/ModernBERT-base IS in the frozen pricing_v15.UNLISTED_BASE_MODELS, so no exact base-model floor applies and the module…
  • openJev Verdict 1.4: ~$0.0028 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Raw Qwen3 0.6B direct logits: ~$0.0034 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • lev-350m: ~$0.0046 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Raw Qwen3 1.7B direct logits: ~$0.0034 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • kev 0.5B: ~$0.0034 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • Jez-small-1.0.0: ~$0.0029 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • openJev Verdict: ~$0.0028 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Quyet-1.0-Small: ~$0.0048 est. per 1,000 decisions. v1.5.8 fast-lane delivery row (combined v1.5.8 preview, not yet published) cost per 1,000 decisions carried (pricing rules unchanged.
  • watt-flash-0.1: ~$0.0023 est. per 1,000 decisions. ESTIMATE (size-class reference; self-hosted open weights, no public tariff).
  • Quyet-1.0-Tiny: ~$0.0024 est. per 1,000 decisions. v1.5.8 fast-lane delivery row (combined v1.5.8 preview, not yet published) cost per 1,000 decisions carried (pricing rules unchanged.
  • Phocinae-Largha-150M-v1: ~$0.0014 est. per 1,000 decisions. ESTIMATE (*est.), size-class reference; self-hosted open weights, no public bookable tariff.
  • Liquid AI d1-omni-600M: ~$0.0079 est. per 1,000 decisions. ESTIMATE (*est.), size-class reference; self-hosted open weights, no bookable listing for this base.
  • Vega 0.8B: ~$0.0116 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • open-jev-deberta-v3-large: ~$0.0056 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Julia-1: ~$0.0022 est. per 1,000 decisions. ESTIMATE (*est.), size-class reference; self-hosted open weights, no public bookable tariff for these weights.
  • verdict-small: ~$0.0009 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Tacet Sonata: ~$0.0023 est. per 1,000 decisions. ESTIMATE (size-class reference; self-hosted open weights, no public tariff).
  • Laya: ~$0.0032 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Mixedbread mxbai-rerank-base-v2: ~$0.0210 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Laya multilingual: ~$0.0039 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Certo v1: ~$0.0013 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • CLM-8B: ~$0.0075 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.
  • BAAI bge-reranker-v2-m3: ~$0.0227 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Alibaba GTE Reranker ModernBERT-base: ~$0.0106 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Mirror: ~$0.0023 est. per 1,000 decisions. v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged.
  • Open Jev JSON Canvas: ~$0.0095 est. per 1,000 decisions. GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour.

Loading context-length views…

Not yet measured on v1.6 · 13 open-weights systems with a dated carried score

A further 1 carried rows are hosted API offerings; they appear here when “Show API offerings” is on and are listed on the API leaderboard.

v1.5 These rows were measured before the 10 s rule (v1.5 method), not re-scored: their scores are not comparable with the re-scored rows above.

Every ranked system of the live v1.5.7 board that is not measured on the v1.6.0 pool keeps its last published score, marked with the release that first published that measurement and that release's publication day. Carried rows are listed separately and never ranked together with v1.6-measured rows. v1.5 protocol (1,624 decisions: 904 open + 720 sealed); scores are on the v1.5 scale and are not comparable with v1.6-measured rows.

Dates: v1.5.0 published 2026-09-28 (12) · v1.5.3 published 2026-09-29 (1) · plus 1 hosted API rows when shown. The date is the publication day of the release that first published the measurement, not a per-model measurement timestamp.

Carried JevBench v1.5.x results, not ranked with v1.6 measurements
SystemMeasured onCapability (v1.5 scale)v1.5 CompositeCost / 1,000Median latency
AutoJev-27B (RTX PRO 6000)v1.5measured on v1.5.0 (2026-09-28)79.719.5 · was #56 on v1.5.7$0.2262estimate0.33 s
Surogate Rune 26B-A4B v3 (RTX PRO 6000)v1.5Not yet measured on the v1.6 pool.measured on v1.5.0 (2026-09-28)79.066.5 · was #19 on v1.5.7$0.0502estimate0.35 s
reflex-27b (Qwen3.8-27B)v1.5measured on v1.5.0 (2026-09-28)74.413.2 · was #68 on v1.5.7$0.2973estimate2.75 s
NInfer Qwen3.8-27B NVFP4 (T=1.5)v1.5measured on v1.5.0 (2026-09-28)73.818.5 · was #59 on v1.5.7$0.2305estimate0.25 s
jeff (Logan Markewich, GLiFormer 400M)v1.5measured on v1.5.0 (2026-09-28)42.40.1 · was #89 on v1.5.7$0.0043estimate7.14 s
Von (wfzyx, Option-Marker 395M)v1.5measured on v1.5.0 (2026-09-28)41.70.0 · was #110 on v1.5.7$0.0038estimate0.92 s
Bosun v3.1 0.6Bv1.5measured on v1.5.3 (2026-09-29)39.42.5 · was #78 on v1.5.7$0.0056estimate4.04 s
GLiNER2.5 multi (Fastino, 287M)v1.5measured on v1.5.0 (2026-09-28)36.02.7 · was #77 on v1.5.7$0.0028estimate1.37 s
GLiNER2.5 small (Fastino, 74M)v1.5measured on v1.5.0 (2026-09-28)32.50.9 · was #85 on v1.5.7$0.0028estimate0.47 s
GLiNER2 large (Fastino)v1.5measured on v1.5.0 (2026-09-28)32.48.4 · was #70 on v1.5.7$0.0056estimate1.88 s
GLiNER2 (Fastino, gliner2.5-base)v1.5measured on v1.5.0 (2026-09-28)24.52.3 · was #79 on v1.5.7$0.0028estimate1.00 s
Needle 3 (Cactus, 2-bit, local CPU)v1.5measured on v1.5.0 (2026-09-28)0.00.0 · was #102 on v1.5.7$0.0191estimate135.51 s
Needle 3, options as tools (post-hoc adapter mode)v1.5measured on v1.5.0 (2026-09-28)0.00.0 · was #103 on v1.5.7$0.0191estimate58.16 s

* Self-hosted rows: the first value is the p50 we measured on our own hardware over loopback (on-GPU, or on-CPU where marked, no network). It is not directly comparable with API latency, which we time end to end from our measuring host, network included. “est.” is the est. incl. network (2x + 150 ms) that the ranking and the latency cap use for every self-hosted row. Model authors can ask us to time a hosted endpoint instead (on request).

How it works

One merged board: how newer draws join it

Each system appears once, with its latest valid measurement; earlier systems are not re-measured. The reference scale is v1.6.1 (draw v1.6.0). Every draw shares the same 300 public items (P, manifest 6b2321f8...). A fresh draw is equated to the reference draw with fixed anchor systems measured on both draws: offset = median over anchors of (metric on the reference S u P) - (metric on the new draw S u P), for Intelligence and Calibration, with a bootstrap 95% interval over anchors and items. A row from that draw is shown as published plus the offset; Capability and Composite follow from the equated axes. Until the anchor runs of a draw are complete its offset is 0 and the draw is marked 'anchor runs pending'.

DrawDrawnReleasesAnchor offset (Intelligence / Calibration)Status
v1.6.0 (reference)2026-10-01v1.6.0, v1.6.10 / 0reference scale · G_med 2.57
Fast-lane draw · v1.6-fastlane-202610092026-10-09v1.6.2, v1.6.3, v1.6.4+0.00 / +0.00anchor runs pending: shown as published
Regular draw · v1.6-regular-20261010-a22026-10-10v1.6.5, v1.6.6, v1.6.7, v1.6.8, v1.6.9+0.00 / +0.00anchor runs pending: shown as published
Second regular draw · v1.6-regular-202610122026-10-12v1.6.10, v1.6.11, v1.6.12, v1.6.13, v1.6.14, v1.6.15, v1.6.16, v1.6.17, v1.6.18, v1.6.20, v1.6.21, v1.6.22, v1.6.23, v1.6.24, v1.6.25, v1.6.26, v1.6.27, v1.6.28, v1.6.29, v1.6.30, v1.6.31, v1.6.32, v1.6.33, v1.6.34, v1.6.35, v1.6.36, v1.6.37, v1.6.38, v1.6.39, v1.6.40, v1.6.42, v1.6.43+0.00 / +0.00anchor runs pending: shown as published
Third regular draw · v1.6-regular-202610272026-10-27v1.7.13, v1.7.14, v1.7.15, v1.7.16, v1.7.17, v1.7.19, v1.7.21, v1.7.22, v1.8.1, v1.8.2, v1.8.3, v1.8.4, v1.8.5, v1.8.6, v1.8.7+0.00 / +0.00anchor runs pending: shown as published
  • Anchor pool: hopper, kev-0.6b, kev-4b, kev-8b, localjev-qwen3.5-4b, malkuth-4b, metask-jev-4b, raw-phi-4-mini, raw-qwen3-1.7b, raw-qwen3-4b-instruct-2507, raw-qwen3-8b, typecastlm. The same 12 anchors answered two sealed draws of v1.6.0 (S and the A2 supplement): median difference -0.29 Intelligence and +1.56 Calibration points (v1.6.1 results, v16.equating_A2).
  • On the shared public items, systems on both new draws score 6 to 13 sealed points below the reference field's sealed-vs-public line (reference residual SD 3.8). Either the new sealed draws are harder or their cohorts are tuned on the public items; only anchor runs can separate the two. Until then their rows are, if anything, understated.
  • Rows from a newer draw keep the gap-penalty reference (G_med) their release was scored with; the row tooltip and tag name the draw and date. Their topic, use-case and language cells use their own draw’s item set and labels, so they are shown on their release page (linked from the draw tag in Release history), not in the radars and language table here.

Open weights and API offerings (board v1.7.48)

Why open weights and APIs are compared in separate groups
  • Fair cost and speed. We run every open-weights model on hardware we rent and operate, so cost and latency compare on the same terms. The GPU cost calculator prices your own setup: own hardware, on-demand or long-term rental.
  • API prices can change. An API price is the vendor's decision. It can be subsidised (for example on top-end GPUs) and raised later, and readers cannot reproduce it.
  • Different fairness needs. A hosted endpoint chooses its own hardware and sees the benchmark inputs, so API offerings are compared with each other.
  • Open-source focus. JevBench exists to make open decision models comparable and reproducible. Every score and the method are identical in every view; the All view ranks both groups together by Capability.
  • This board ranks open-weights systems: published weights that we ran ourselves, on GPU or CPU machines we rent and operate. A system we measured through an endpoint we do not run (vendor API, author-hosted or third-party-hosted endpoint, for example Qwen3.8 27B via Chutes) is an API offering, even when its base weights are open; API offerings are ranked on the API leaderboard.
  • Why separate boards: open weights can be compared on equal hosting terms, while an API price is a vendor decision that can be subsidised or raised later and is not reproducible by readers. Every score and measurement is the same on both boards; only the set of ranked rows differs, and ranks are the published order filtered to that set.
  • Jev 1.13.0 is a hosted API. It stays on the open-weights board as the reference row (it defines the Jev-class cost and latency caps) and is not ranked there; it is ranked on the API leaderboard. Since v1.6.30 the board also shows Jev 1.14.0, TypeSafe's current version, as a second unranked reference row for comparison (display only; it is ranked on the API leaderboard and does not define any cap).
  • Official cost basis is unchanged: the Cost axis keeps each row's documented reference price (see the cost notes below; APIs with a known base model are priced at the developer's own list price). The base-model reference price only applies to open-weights rows; API offerings are ranked at their own list price. The GPU cost calculator above is a What-If for your own hosting and never changes a score or rank.
  • Full API re-run (A4, v1.7.7). Most API offerings were measured on 6 Oct 2026 on a fresh sealed API set A4 (300 never-used sealed items) plus the same 300 public items every system answers, and equated to the v1.6.1 S ∪ P scale with the published A2/A3 supplement method (offset +2.57 Intelligence, +3.35 Calibration; pool of 8 ranked self-hosted systems re-run on A4 ∪ P). Jev, Sage, wity-1 and Fastino GLiNER-2.5-Decide keep their full-set S ∪ P scores; Liquid AI d1 answered the full S ∪ P set on 6 Oct 2026 (v1.7.8) and is scored like them, without equating. The pool is mid-strength, so for the strongest LLM rows the offset is an extrapolation (likely within ±2 Intelligence points; the 95% intervals include it). Cost rule (v1.6.1, generalised in v1.6.30): On every draw, very long items that a ranked row refuses because they exceed its context length are left out of cost for all rows: cost per 1,000 decisions is computed on the same remaining items for every row. A refused item still counts as a wrong answer for Intelligence. On A4 ∪ P this left out nine items of about 77,000–82,000 input tokens. First published as the v1.6.1 cost rule (5 Oct 2026; 23 items on that draw). From v1.6.30 it applies to draw v1.6-regular-20261012: the 15 items the Jev reference refused with max_tokens_exceeded are left out and 1,485 remain. Rows of draw v1.6-regular-20261010-a2 still show the all-items cost they were published with until a following release re-costs them under this rule. Rows without a public tariff keep their documented v1.5 cost estimate. A4 is now retired for everyone. OpenAI Decisions (gpt-6-luna) ran later on the fresh sealed API set A5 ∪ P (600 items; A4 had already been sent to that provider) and is equated the same way with A5's own offsets (+5.78 Intelligence, +6.15 Calibration, same pool).
  • Language and use-case cells (v1.7.12). These breakdowns add two sealed supplements (354 language items, 33 use-case items, drawn and reviewed on 6 Oct 2026), so every language and use case has at least 30 items. Headline scores are unchanged; rows not yet run on the supplements are tagged S+P in the language table.
  • Every API offering we reached is measured on v1.6.1: either on the full 1,500-item set or on A4 ∪ P (600 items, equated). No preliminary public-set rows remain.

Revision history

  • v1.7.48 (2026-10-22): Method v1.7 (JevBench v1.7.0): a decision over 10 seconds counts as a wrong answer and as zero calibration for that item (measured time); Speed uses all decisions; every row shows its measured p95 and how many decisions were over 10 s. All rows were re-scored from their recorded per-decision times; rows recorded on a loaded CPU host are listed as re-measure pending with their v1.6 values. The v1.6 values stay visible on every row.
  • v1.7.47 (2026-10-21): Announcement only: JevBench v1.7.0 (not before 24 October 2026) adds a 10-second limit per decision; decisions over 10 seconds count as failed, and Speed uses all decisions. No score or rank changed in this revision.
  • v1.7.46 (2026-10-11): Display: the measured value of self-hosted rows is now tagged “on-GPU, no network” (or CPU) and marked as not directly comparable with API latency, which is timed end to end from our measuring host. The method notes describe a planned calibration of the 2× + 150 ms assumption against real end-to-end timings. No score or rank changed.
  • v1.7.45 (2026-10-11): Display: self-hosted rows now show the measured p50 next to the network-including estimate (2× measured + 150 ms), which the ranking continues to use. No score or rank changed.
  • v1.7.44 (2026-10-10): Display only: the Pareto charts get the same “Show models that don’t qualify as Jev-class” switch as the scatter charts (off by default, so the frontier is drawn over Jev-class systems), expand to fullscreen, and share the highlighted system with the scatter charts (hover or tap one system and it is marked in all four charts). No score or rank changed.
  • v1.7.43 (2026-10-10): Display only: a Pareto frontier section after the two scatter charts on both boards — Capability against cost per 1,000 decisions and against median latency (log axes; list price and measured endpoint latency on the API board). The red line joins the systems no other system beats on both axes. No score or rank changed.
  • v1.7.42 (2026-10-10): One merged board. The version tabs are gone: /jev-models is always the current board with every system’s latest measurement, and releases are listed under Release history (old version pages keep their URLs). The 16 systems measured on the two fresh draws after v1.6.1 (fast-lane draw v1.6.2–v1.6.4, regular draw v1.6.5–v1.6.7) join the boards with their published scores and a draw tag; each draw is put on the v1.6.1 scale by an anchor offset that stays 0 until the anchor systems have run on it (see How it works). A filter on top switches between Open weights, API and All (both groups, ranked by Capability). Header text is shortened; the details moved into How it works. Jeff 1.0 Large (fast-lane draw) enters the open-weights Capability top 5 at #2.
  • v1.7.41 (2026-10-10): Needle 3 now has language and category cells on the current pools. It keeps its published v1.5 headline. With the benchmark owner's approval its closed engine was sent all supplement items (L1, L2, L3) once more on an isolated machine, plus 128 never-asked S+P items chosen by their use-case label to complete the last radar spoke (these also count in the other cells, so its S+P part is not a random sample); together with an earlier partial run both radars reach at least 30 answered items per spoke (52 per subject topic, 43 per use case). The engine abstains on about 30 % of items and returns invalid UTF-8 on about 7 %; both count as unanswered, so 22 languages stay below the 60-item target (19 to 59 answered items). Its answers to the same public items differ between machines on about 10 %. Its chance-corrected competence is below zero in every cell, shown as 0. The v1.7.40 note for Needle 3 (options as tools) is corrected the same way. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3. Headline scores, Capability, Composite, ranks and pricing are unchanged.
  • v1.7.40 (2026-10-10): Needle 3, options as tools, now has language and category cells on the current pools. It keeps its published v1.5 headline. With the benchmark owner's approval its closed engine ran once more on an isolated machine; the run covered most of L3 and half of L1 before the machine's time limit and is combined with an earlier partial run. Both radars are complete (at least 58 per subject topic and 43 per use case). The engine abstains on about 9 % of items and returns invalid UTF-8 on about 7 % (mostly non-Latin scripts); both count as unanswered, so 18 languages stay below the 60-item target (24 to 59 answered items; columns under 30 carry the low-n mark). Its chance-corrected competence is below zero in every cell, shown as 0. ClassOne Gemma 4 E2B keeps its exception with a corrected reason: its weights were identical, but its server gives different answers after each restart. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3. Headline scores, Capability, Composite, ranks and pricing are unchanged.
  • v1.7.39 (2026-10-10): GLiNER2 large now has language and category cells on the current pools: at least 62 completed responses in every language, 98 in every subject topic and 80 in every use case. It keeps its published v1.5 headline; the cells come from a new run with the original pinned model on an isolated machine. GLiNER2 gained 180 more answered items that had been reserved earlier but never sent: its languages now rest on 50 to 59 answered items where they are below the 60-item target (17 of 23), and its radars on at least 69 per topic and 56 per use case. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3. Headline scores, Capability, Composite, ranks and pricing are unchanged.
  • v1.7.38 (2026-10-10): Three more historical rows now have language and category cells on the current pools: reflex-27b, Surogate Rune 26B-A4B v3 (RTX PRO 6000) and jeff. They keep their published v1.5 headlines. reflex-27b completed its partly answered S+P set and the supplements; Surogate Rune adds the supplements to its 6 October S+P run; jeff (CPU) answered all pools. Their repeated answers to the 300 public items agree 100 %. Each row has at least 65 completed responses in every language, 109 in every subject topic and 83 in every use case. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3. Headline scores, Capability, Composite, ranks and pricing are unchanged.
  • v1.7.37 (2026-10-10): Three more historical rows now have language and category cells on the current pools: Von, GLiNER2 and GLiNER2.5 multi. They keep their published v1.5 headlines; the cells come from new runs with the original pinned model and setup on isolated machines. Von and GLiNER2.5 multi reach at least 61 completed responses in every language, 100 in every subject topic and 80 in every use case. GLiNER2 fills both radars (at least 57 per topic, 45 per use case) and every language column, but about 1,000 of its issued items were lost to out-of-memory kills and a machine shutdown; they stay counted as spent, so 22 languages rest on 42 to 59 answered items, below the 60-item target. Three rows that cannot be measured now give the concrete reason: Aplomb 1 (pinned revision deleted upstream), ClassOne Gemma 4 E2B (re-run did not reproduce its original answers) and swanOne (weights return HTTP 401). L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3. Headline scores, Capability, Composite, ranks and pricing are unchanged.
  • v1.7.36 (2026-10-10): Nine more rows now cover S+P+L1+L2+L3 in the language table and both radars: Diffusion Jev, SPX-CD-Omni, Seb-9B, APUS-OpenJev-v1-9B, APUS-OpenJev-v1-35B-A3B, Jobe Qwen3.5-4B, AutoJev-27B (RTX PRO 6000), NInfer Qwen3.8-27B NVFP4 (T=1.5) and OpenJev (DiffusionGemma 26B-A4B NVFP4, razorback16). The last four are historical rows: they keep their published v1.5 headlines, and their language and category cells come from a new run on the current pools with the original pinned model and serving setup. Diffusion Jev, SPX-CD-Omni, Seb-9B and both APUS rows match their original public answers on at least 97 %. The razorback16 OpenJev model samples stochastically, so its repeated passes agree on 86 to 89 % of the public items; its note says so. Every row has at least 65 completed responses in every language, 109 in every subject topic and 83 in every use case. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3. Headline scores, Capability, Composite, ranks and pricing are unchanged.
  • v1.7.35 (2026-10-10): Seven more rows now cover S+P+L1+L2+L3 in the language table and both radars: René-1 31B FP8, decisio v0.8.0 on gemma-4-12B-it, Hopper 12B trained, decider-12b v2, decider-12b v1, Bobcat Flash 1.2 and Mica v0.1 4B. Each answered all 1,505 supplemental items with its original pinned model and serving setup on an isolated machine; answers to the 300 public items match the original runs (at least 99 %). Mica keeps its published v1.5 headline; its language and category cells come from a new run on the current pools. Every row has at least 65 completed responses in every language, 109 in every subject topic and 83 in every use case. The language table now lists the mixed-language group first. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3. Headline scores, Capability, Composite, ranks and pricing are unchanged.
  • v1.7.34 (2026-10-10): Seven more self-hosted rows now cover S+P+L1+L2+L3 in the language table and both radars: TypeCastLM 1.4.0, CoCo-Decision-4B, ClassOne Qwen 3.5 9B, EXAONE-4.0-1.2B-JEV v0.3, jul fast, WaterSheep and Tacet Sonata. Each answered all 1,505 supplemental items with its original pinned model and serving setup on an isolated machine; their answers to the 300 public items match the original runs (at least 97 %). Every row has at least 65 completed responses in every language, 109 in every subject topic and 83 in every use case. Long-input refusals stay scored as before. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3. Headline scores, Capability, Composite, ranks and pricing are unchanged.
  • v1.7.33 (2026-10-10): Messier One v0.2 now covers S+P+L1+L2+L3, with all 1,505 supplemental items answered natively on the original image with the network disabled: at least 65 completed responses in every language, 109 in every subject topic and 83 in every use case. The original P300 answers recur in each supplement and count once; the original S+P measurement had no HTTP errors. L3 items were reviewed by an OpenAI model, so cells based on L3 carry that exposure; headline scores do not use L3. Headline scores, Capability, Composite, ranks and original row pricing are unchanged.
  • v1.7.31 (2026-10-10): Microsoft-Decision-1 joins the hosted API board after a native Azure Foundry measurement on 1,200 sealed plus 300 public observations: Capability 70.78, Calibration 84.40 and Composite A 69.11. Its 21 context-limit failures remain in the scored denominator. Native median latency is 459 ms on the 124-request speed subset. This new full-set O1S run uses its completed-field median gap reference of 7.088 score points; historical rows retain their own dates, pools and references rather than being remeasured or converted. The existing top five on Composite A and capped Capability are unchanged. All 50 topic/use-case/language cells are recorded; thin cells stay suppressed or table-only, and qualified supplemental coverage remains pending.
  • v1.7.30 (2026-10-10): decisio v0.8.0 on gemma-4-31B-it now covers S+P+L1+L2+L3, with all 1,505 supplemental items answered natively: at least 65 completed responses in every language, 109 in every subject topic and 83 in every use case. The original P300 answers recur in each supplement and count once. The original 23 HTTP 500 errors remain operational non-answers and are not imputed. L3 items were reviewed by an OpenAI model, so its cells based on L3 carry that exposure; this row is the original locally served Gemma offering. Headline scores, Composite, ranks and original row pricing are unchanged.
  • v1.7.29 (2026-10-09): Mercury Decide now has at least 64 completed observations in every language, 109 in every subject topic and 83 in every use case, using S+P+L1+L2+L3. L3 retains 1,416 records on its full 1,418-item denominator, including original refusals and operational failures; two spent failures remain missing. Paid supplemental decisions use the currently declared Inception 2026-09-30 offering and the original native protocol. Public top labels agree on 96.63%, with differing probabilities and unproven historical weights/calibrator equality; the row coverage note gives the comparison. Headline scores, Composite, ranks and original row pricing are unchanged.
  • v1.7.28 (2026-10-09): GLiNER 2.5 Small now covers P300 + L1 + L2 + L3 with 1,802 completed native responses: at least 60 in each of 23 languages, 78 in every subject topic and 52 in every use case. Its original offline CPU/fp32 profile is unchanged. Three earlier public-item OOM outcomes remain missing and contribute no completed coverage; no requests were replayed. These cells exclude historical S answers and stay outside headline scores. Headline scores, Capability, Composite and every rank are unchanged.
  • v1.7.27 (2026-10-09): OpenJev DeBERTa v3 Large adds 60 new L5 items: 20 each in Arabic, Hindi and Greek. It now has at least 63 completed responses in every language (Arabic 70, Hindi 73, Greek 69). The original native context was checked before keeper selection; all 60 requests succeeded, and earlier failures remain scored. Other rows and category views retain their existing measurements. Headline scores, Capability, Composite and every rank are unchanged.
  • v1.7.26 (2026-10-08): BB-Qwen3.5-4B-LoRA now covers S+P+L1+L2+L3, with at least 65 completed responses in each of 23 languages and at least 83 on every capability and use-case spoke. Two DNS failures remain scored and do not count toward completed coverage. Its reference cost stays unknown; these view values do not impute a price. OpenJev DeBERTa v3 Large now includes its completed L1/L2 runs; Arabic (50), Hindi (53) and Greek (49) still need more completed responses. Headline scores, Capability, Composite and every rank are unchanged.
  • v1.7.25 (2026-10-08): Watt Flash 0.1 and OpenJev Verdict completed their language supplements. Both now cover S+P+L1+L2+L3, with at least 65 completed responses in each of 23 languages and at least 83 in each capability and use-case category. Other rows retain their existing measurements, including Fastino’s Hindi L4 supplement. Headline scores, Capability, Composite and every rank are unchanged.
  • v1.7.24 (2026-10-08): Fastino GLiNER 2.5 Decide now has 69 completed Hindi responses after a new 20-item Hindi supplement. Seventeen requests succeeded; three HTTP 400 failures remain scored and do not count toward completed coverage. L4 is used only for this offering’s Hindi language cell, with OpenAI author and independent Anthropic reviewer provenance disclosed. The mghafiri Qwen3.5 0.8B row also completed L3, expanding its language and category cells. Headline scores, Capability, Composite and every rank are unchanged.
  • v1.7.23 (2026-10-08): The Languages table includes every listed model, including historical and catalogue entries. These rows use their stored language measurements where available and show pending cells where measurements are unfinished. Historical headline scores are not used as language values. Headline scores, Capability, Composite and every rank are unchanged.
  • v1.7.22 (2026-10-08): Languages and radars now include completed L1/L2/L3 runs for SPX CD Flash, SPX CD Pro and Decisio Gemma 4 12B v0.9. Coverage counts completed supported responses and recorded input refusals; authentication, rate-limit, service and transport errors do not satisfy reporting thresholds. Competence keeps its original scored observations. 2 listed rows still have incomplete radar coverage and show the reason. Every listed row remains in the completion check. Headline scores, Capability, Composite and every rank are unchanged.
  • v1.7.21 (2026-10-08): Languages and radars: more rows completed their L1/L2/L3 supplement runs, among them Instinct (its API is back), Surogate Rune 26B-A4B v3, Xor 26B-A4B, JADE, Jebadiah 27B, Kev 27B, Decision 2.0 Vega 27B and the self-hosted SimpleJev Qwen3.8-27B. 2 rows still show a reason instead of a full radar: newly listed rows whose supplement runs are not done yet, rows not measured on this item pool, and one model whose pinned weights could not be accessed. Headline scores, Capability, Composite and every rank are unchanged.
  • v1.7.20 (2026-10-08): Correction and progress for the language and radar views. The self-hosted row SimpleJev Qwen3.8-27B wrongly counted the hosted SimpleJev API row's L1/L2/L3 supplement answers in its language and category cells (v1.7.18 and v1.7.19), because both share a file name. Its cells now use only its own S + P answers until its own supplement runs finish, and the cell builder no longer lets a hosted-API run count for a self-hosted row. More open-weights rows completed their L1/L2/L3 runs. 2 rows still show a reason instead of a full radar. Headline scores, Capability, Composite and every rank are unchanged.
  • v1.7.19 (2026-10-08): Languages and radars: the open-weights rows were run on the L3 supplement overnight, and most rows that still lacked the earlier L1/L2 supplements got them too. Their language cells and topic/use-case radars now count every item they answered (S + P + L1 + L2 + L3), and Qwen3.8 27B (Chutes) finished its L3 run. L3 items were reviewed by an OpenAI model, so the L3-based language and category cells of OpenAI rows carry that exposure; a note next to those rows says so, and headline scores do not use L3. 2 rows still show a reason instead of a full radar, for example runs still in progress, a provider that is down, withdrawn weights or rows not measured on this pool. Headline scores, Capability, Composite and every rank are unchanged.
  • v1.7.18 (2026-10-08): Languages and radars for every row. A new sealed supplement, L3 (1,118 items drawn on 7 Oct 2026: items written natively in 21 languages plus a mixed-language group, and English items for thin radar categories such as everyday language, safety and the "other" use case), gives every language at least 60 items on P ∪ L1 ∪ L2 ∪ L3 (before: as few as 4). The hosted API rows answered it: their language cells and their use-case and topic radars now rest on about 2,100 answered items (3,000 for rows that answered the full sealed set), with at least 69 on every spoke. Open-weights rows are being run on it in the current GPU wave and update as they finish. 2 rows that do not reach 30 items on every spoke yet show the reason next to the row. Headline scores, Capability, Composite and every rank are unchanged.
  • v1.7.17 (2026-10-07): Languages now includes every measured row on both boards, with wrappers listed below the models. Self-hosted rows use S 1,200 + P 300 + L1 (354 items) + L2 (33 items) + L3 (1,118 items), where answered. API rows re-run on A4/A5 use their 300/300-item sealed subset instead of S. Every row’s tag lists the pools it actually answered. L3 is a sealed language supplement (1,118 items), drawn 2026-10-07. Every language has at least 60 items in P ∪ L1 ∪ L2 ∪ L3. Header counts show the full S + P + L1 + L2 + L3 pool. Headline scores are unchanged; these raw cells are unequated, scored for language/category views only and outside the Composite. L3 items were written natively by Claude Sonnet 5.5, each solved blind and language-checked by GPT-6.1 Sol, with gold kept only when both agree or a second review confirms; no gold comes from Jev or any measured API. L3 is API-facing by design and is excluded from future headline draws. Rows with unfinished runs retain their actual coverage tags. L3 items were reviewed by an OpenAI model, so the L3-based language and category cells of OpenAI rows (OpenAI Decisions, GPT-6 Luna, GPT-5.6 Luna) carry that exposure; headline scores do not use L3. C1 adds English items for thin radar categories (everyday language, safety and the other use case). L4 adds 20 new Hindi items for Fastino GLiNER 2.5 Decide only. They were authored with OpenAI and independently solved and language-reviewed with Anthropic before sealing. Seventeen L4 requests succeeded; three HTTP 400 failures remain in the scored observations and do not count as completed coverage. L4 is outside headline and category scores and the common header counts. L5 adds 60 new items for OpenJev DeBERTa v3 Large only: 20 each in Arabic, Hindi and Greek. They were authored with OpenAI and independently solved and language-reviewed with Anthropic before sealing, then checked against the original 512-token native context before keeper selection. All 60 native requests succeeded. Previous failures remain scored and do not count as completed coverage. L5 is outside headline and category scores and the common header counts. 19 API rows re-run on A4/A5 have answered L1/L2 supplements; their category radars count those items too. L3 coverage will follow when measured.
  • v1.7.16 (2026-10-07): Compare view, API board: subject-topic and use-case radars for OpenAI Decisions, the 17 API offerings re-run on A4 ∪ P and the classifier.dev wrapper. Their 600 new sealed items were labelled with the same recipe as every other item (Winnow-12B Q8 on our own GPU pod; uc1 items keep their authoring use case); values are raw and rest on 600 items (300 sealed), so more categories fall under the 30-item spoke minimum and are listed below the radar. Each of these rows also gets a breakdown section on its own page. Every live row now has all breakdown views, and a release check keeps it that way. No score, axis, cost or rank changed.
  • v1.7.15 (2026-10-07): Compare view, API board: OpenAI Decisions, the 17 API offerings re-run on A4 ∪ P and the classifier.dev wrapper now show competence per request type and per tier on the open and sealed items they answered (300 open + 300 sealed: fewer sealed items than the full set, so wider uncertainty; the view note says so), and Liquid AI d1 shows its subject-topic and use-case radars. Values are raw, computed from the stored per-item results with the same scorer as every other row. Subject-topic and use-case radars for the 600-item rows follow once their 600 new sealed items are labelled. No score, axis, cost or rank changed.
  • v1.7.14 (2026-10-07): Open-weights board: H2O-Lightning-4B v1.1 (H2O.ai) and Kahn1 4B (Okura66), two Qwen3.5-4B fine-tunes submitted as JevBench add-requests, measured on the same sealed v1.6 pool with the same scorer, G_med and cost basis as every other row (addendum a8 of v1.6.1). Both are priced with labelled cost estimates like every other Qwen3.5-4B row; Kahn1's server reports no token usage, so its estimate is our own count of its input tokens (three passes per decision, its server default). Both rank inside the Jev-class caps: H2O-Lightning-4B at #3 of the Capability Score and #1 of the Composite, Kahn1 at #23 of the Capability Score. No published score, axis or cost changed.
  • v1.7.13 (2026-10-07): Open-weights board: Surogate Rune 26B-A4B v3 (measured on the v1.6 pool for the first time; its dated v1.5 row stays in the carried list) and Xor 26B-A4B (Juspay), both from the top 20 of the community Jev Decision Index on Hugging Face, measured on the same sealed v1.6 pool with the same scorer, G_med and cost basis as every other row (addendum a7 of v1.6.1). Both rank inside the Jev-class caps, at #3 and #4 of the Capability Score. No published score, axis or cost changed.
  • v1.7.12 (2026-10-07): Languages and use cases: the per-language and per-use-case cells now also count two sealed supplements, a language supplement L1 (354 items) and a use-case supplement L2 (33 items), drawn and reviewed on 6 Oct 2026. Every one of the 22 languages (plus the mixed-language group) and every one of the 20 use cases now has at least 30 items (before: as few as 5). 82 rows answered both supplements; rows not run on them yet keep their S + P cells and are tagged S+P (or S+P+L1). Headline scores, Capability, Composite and every rank are unchanged on both boards.
  • v1.7.11 (2026-10-07): Open-weights board: 12 new self-hosted Jev-compatible rows from the top 20 of the community Jev Decision Index on Hugging Face (Perplexity Decider v1.1 27B, torchcast-decision-27b, Kev 27B, decider chat on Gemma-4-31B-it, GEV-26B-Decide, Decision 2.0 Vega 27B, JEV-27B, Jebadiah 27B, JADE, Bespoke Nimble 9B v3, SimpleJev Qwen3.8-27B self-hosted, JPT-35B-A3B), measured on the same sealed v1.6 pool with the same scorer, G_med and cost basis as every other row (addendum a6 of v1.6.1). No published score, axis or cost changed; ranks move only where a new row places above. Rows priced above the Jev-class cost or latency caps are listed under the limits section, not in the Capability ranking.
  • v1.7.10 (2026-10-06): API board: OpenAI Decisions (POST /v1/decisions with gpt-6-luna) replaces its preliminary public-set row with its official result. It answered a fresh sealed API set A5 (300 never-used sealed items, since A4 had already been sent to OpenAI) plus the 300 public items (598 of 600 answered) and is equated to the v1.6.1 scale with the same method and pool as the A4 rows (offsets +5.78 Intelligence, +6.15 Calibration). It is ranked on the API board: Composite 62.5 (95% interval 51.6–64.6), Capability 73.5; list price USD 0.10 per 1M input tokens as of 6 Oct 2026 (launch day), re-checked at each revision. It enters the API Composite top 5 at #5, statistically tied with Instinct (62.2, now #6), and the Jev-class Capability top 5 at #4 (Instinct moves to #5, Vansa-3.4 to #6). Every other score is unchanged; the open-weights board is unchanged.
  • v1.7.9 (2026-10-06): API board: the OpenAI Decisions API (POST /v1/decisions with gpt-6-luna, opened on 6 Oct 2026) is added as a preliminary, unranked row from the 300 public v1.6 items (298 answered; Capability 68.8, Composite 37.2 on that set; list price USD 0.10 per 1M input tokens). Its official sealed run waits for the next fresh sealed API draw. No score or rank changed on either board.
  • v1.7.8 (2026-10-06): API board: Liquid AI d1 is added as a ranked API offering. It answered the full v1.6.1 set (1,200 sealed + 300 public items, all 1,500 answered) on 6 Oct 2026 and is scored exactly like the other full-set API rows, without equating (Composite 73.0, Capability 74.4; tariff USD 0.04 per 1M input tokens). It enters the API board top 5 in Composite and in Jev-class Capability. Every other score is unchanged; the open-weights board is unchanged.
  • v1.7.7 (2026-10-06): API board: 17 API offerings re-measured in full on a fresh sealed API set (A4, 300 never-used sealed items plus the 300 public items, 6 Oct 2026) and equated to the v1.6.1 scale with the published A2/A3 supplement method. They replace the hatched preliminary rows and are ranked on the API board: Instinct and Vansa-3.4 enter its Composite top 5, and Instinct, Vansa-3.4 and Instinct Dual 4B join Sage and Jev in its Jev-class Capability top 5; GPT-6 Luna, Qwen3.8 27B, GPT-6 Luna low, DeepSeek Flash and GPT-5.6 Luna have the highest raw Capability (98.4, 98.2, 97.8, 97.4 and 94.5) but sit outside the Jev-class caps. Qwen3.8 27B (Chutes) is ranked as well (Composite 0.0: no published tariff, so its documented USD 2.18 estimate per 1,000 decisions gives Cost 0). Instinct and Instinct Dual 4B are now treated as production APIs (public tariff), so no demo-endpoint latency adjustment applies to them. classifier.dev (fast tier) is listed as a wrapper, not ranked. No preliminary rows remain. The open-weights board is unchanged.
  • v1.7.6 (2026-10-06): API board: Vansa-3.4 and Instinct Dual 4B now have v1.6 public-set figures and appear as preliminary rows; Autoloops stays pending while its full run on the fresh sealed set is in progress. Very long items that exceed a provider’s context window count as wrong without ending the run. No rank changed.
  • v1.7.5 (2026-10-06): API board: every API offering with a v1.6 public-set figure is drawn in the Composite and Capability charts as a hatched, unranked “preliminary” row at its score position (public set of 300 items; full sealed re-evaluation running); offerings without any v1.6 figure are greyed “pending” rows. Adds Qwen3.8 27B (Chutes) and Fastino GLiDE to the public-set table. The four ranked rows and every rank are unchanged; the open-weights board is unchanged.
  • v1.7.4 (2026-10-06): API board: every reachable API offering that was not yet re-measured on v1.6 now has a dated public-set figure (the 300 public v1.6 items, no sealed items), shown next to its older v1.5 score with Jev and other ranked APIs on the same 300 items as anchors. Not ranked and not comparable with the 1,500-item headline; no score or rank of either board changed.
  • v1.7.3 (2026-10-06): Display only, no score or rank changed. The API board ranks every offering at its own list price and no longer shows a base-model reference price (that comparison only matters against open weights). The Jev reference row and, with “Show API offerings” on, every API offering now sit at their score position in each ranking section (Composite and Capability) instead of the folded end of the list.
  • v1.7.2 (2026-10-06): Display only. The API roster says why carried API rows were not re-run on v1.6 yet (exposure cadence, retired v1.6.0 sealed set) and that no v1.5 to v1.6 conversion is applied.
  • v1.7.1 (2026-10-06): Display only, no score or rank changed. Headings name the board (open weights / API offerings); the API board leads with the Composite Score and lists every API offering we measured, including mode variants, carried v1.5.x rows and wrappers; the Jev reference row reads “Not ranked, only shown as a reference to compare with” on the open-weights board; long base-model notes became numbered footnotes under the ranking; a short expandable note explains the split.
  • v1.7.0 (2026-10-06): Leaderboard split. /jev-models ranks open-weights systems we ran on our own hardware, with Jev 1.13.0 as an unranked reference row and a “Show API offerings” switch; hosted API offerings are ranked on the new /jev-models/api board. No score was recomputed: ranks are the published order filtered to each board. Adds the GPU cost What-If.
  • v1.6.1 (2026-10-05): Hosted APIs (fastino-gliner-2-5-decide, jev-1.13.0, sage-1.3.0, wity-1, wity-1-always, wity-1-off) answer the full 1,500-item set instead of the 600-item API subset and are no longer equated; self-hosted rows unchanged; whole v1.6.0 sealed draw retired. Cost per 1,000 decisions is measured on one common item set (all items except the 23 outside the Jev reference input range) for every token-priced row.
  • v1.6.0 (2026-10-05): Rotating item draw (1,200 sealed + 300 public); hosted APIs on API subsets equated to the self-hosted scale; /jev-models/v1.6.0 stays available unchanged.
  • Earlier releases: v1.5.7, v1.5.6, v1.5.5 and the historical boards below.

System types (colours)

  • Jev — reference (TypeSafe, closed) — The closed system JevBench is named after, shown as the reference.
  • Closed API (weights not public) — Available through a hosted API; the weights cannot be downloaded.
  • Open weights · LLM decoder — An autoregressive language model with public weights, including fine-tunes, merges and Jev rebuilds.
  • Open weights · diffusion LM — A language model that generates by iterative denoising instead of token by token.
  • Open weights · encoder / classifier — BERT-style encoders, NLI zero-shot classifiers and GLiNER-type models.
  • Open weights · reranker — A cross-encoder or LLM reranker that scores options against the input.
  • Base model control (no decision fine-tune, raw logits) — An official open checkpoint without any decision fine-tune, read out from raw logits, used as a floor.
  • System (router / cascade / ensemble) — Several models combined at inference time.

Colours show architecture only; they never change a score or rank. Each class is assigned from cited evidence (config.json, model card, provider docs, or our own run record for hosted APIs).

Self-hosted rows use S 1,200 + P 300 + L1 (354 items) + L2 (33 items) + L3 (1,118 items), where answered. API rows re-run on A4/A5 use their 300/300-item sealed subset instead of S. Every row’s tag lists the pools it actually answered. L3 is a sealed language supplement (1,118 items), drawn 2026-10-07. Every language has at least 60 items in P ∪ L1 ∪ L2 ∪ L3. Header counts show the full S + P + L1 + L2 + L3 pool. Headline scores are unchanged; these raw cells are unequated, scored for language/category views only and outside the Composite. L3 items were written natively by Claude Sonnet 5.5, each solved blind and language-checked by GPT-6.1 Sol, with gold kept only when both agree or a second review confirms; no gold comes from Jev or any measured API. L3 is API-facing by design and is excluded from future headline draws. Rows with unfinished runs retain their actual coverage tags. L3 items were reviewed by an OpenAI model, so the L3-based language and category cells of OpenAI rows (OpenAI Decisions, GPT-6 Luna, GPT-5.6 Luna) carry that exposure; headline scores do not use L3. C1 adds English items for thin radar categories (everyday language, safety and the other use case). L4 adds 20 new Hindi items for Fastino GLiNER 2.5 Decide only. They were authored with OpenAI and independently solved and language-reviewed with Anthropic before sealing. Seventeen L4 requests succeeded; three HTTP 400 failures remain in the scored observations and do not count as completed coverage. L4 is outside headline and category scores and the common header counts. L5 adds 60 new items for OpenJev DeBERTa v3 Large only: 20 each in Arabic, Hindi and Greek. They were authored with OpenAI and independently solved and language-reviewed with Anthropic before sealing, then checked against the original 512-token native context before keeper selection. All 60 native requests succeeded. Previous failures remain scored and do not count as completed coverage. L5 is outside headline and category scores and the common header counts.

Per-system method lists on this board cover the open-weights systems and the two Jev reference rows; the hosted API offerings' lists are on the API leaderboard.

Method amendments

  • 5 Oct 2026, v1.6.1 (published rule): hosted/API systems now answer the same full item set as self-hosted systems (1,200 sealed + 300 public). Their earlier answers on the API subsets were reused and only the missing sealed items were sent; every item was logged in the exposure ledger before it was sent. API rows are no longer equated. Consequence: every v1.6.0 sealed item has now been seen by at least two providers, so the whole v1.6.0 sealed draw is retired for future releases; v1.6.2 onward draws fresh items from the reserve.
  • 5 Oct 2026, v1.6.1 cost rule (published rule): cost per 1,000 decisions is now measured on one common item set for every system: all items except the 23 very long items that lie outside the Jev reference's accepted input range. Before, a system that processed those items paid for their tokens while a system that refused them did not. Refusals still count against Intelligence exactly as before. Hosted APIs with a per-token tariff are priced from their measured tokens on this common set; carried per-decision prices of self-hosted systems are unchanged. Effect: Sage 1.3.0 USD 0.0247 per 1,000 decisions on the common set (0.0650 on all 1,500 items); Jev 1.13.0 0.0323 (unchanged).

Cost rule, generalised in v1.6.30: On every draw, very long items that a ranked row refuses because they exceed its context length are left out of cost for all rows: cost per 1,000 decisions is computed on the same remaining items for every row. A refused item still counts as a wrong answer for Intelligence. First published as the v1.6.1 cost rule (5 Oct 2026; 23 items on that draw). From v1.6.30 it applies to draw v1.6-regular-20261012: the 15 items the Jev reference refused with max_tokens_exceeded are left out and 1,485 remain. Rows of draw v1.6-regular-20261010-a2 still show the all-items cost they were published with until a following release re-costs them under this rule. The 23 items above are the set of the v1.6.1 draw.

The v1.6.1 amendments above say hosted APIs are no longer equated. Exception since v1.7.7: the 17 API rows re-run on A4 ∪ P (600 items) are equated (+2.57 Intelligence, +3.35 Calibration); see “Full API re-run (A4, v1.7.7)” under Open weights and API offerings. Full-set API rows (Jev, Sage, wity-1, Fastino GLiNER-2.5-Decide, d1 (Liquid AI)) are not equated. OpenAI Decisions (gpt-6-luna) ran later on the fresh sealed API set A5 ∪ P (600 items; A4 had already been sent to that provider) and is equated the same way with A5's own offsets (+5.78 Intelligence, +6.15 Calibration, same pool).

Rotating item sets

Each release draws fresh sealed decisions from a larger reserve. Self-hosted open-weights models (run offline on our own GPU pods or Sandy) answer S and P; externally hosted models answer the same full set (S and P) as the self-hosted systems since v1.6.1, so hosted APIs are no longer equated.

SetItemsChoiceNoulScoreAnswered by
S · Sealed release draw1,200600300300self-hosted systems (run offline on our own GPU pods or Sandy) and, from v1.6.1, externally hosted APIs
A · API subset (part of S; used by the v1.6.0 hosted-API measurements)3001507575every system (inside S since v1.6.1); retired for future draws
P · Public set3001507575every system

Not in the table above (frozen release data): API set A4 · 300 never-used sealed items, answered together with P by the 18 API endpoints re-run on 6 Oct 2026 (the 17 ranked offerings and the classifier.dev wrapper); retired after that run.

  • Self-hosted systems: 1,500 items (S 1,200 + P 300). Hosted APIs: 1,500 items (S 1,200 + P 300), the same as self-hosted systems; A (300) is the part of S that earlier API measurements used. Exception since v1.7.7: the 17 API rows re-run on A4 ∪ P (600 items) are equated (+2.57 Intelligence, +3.35 Calibration); see “Full API re-run (A4, v1.7.7)” under Open weights and API offerings. Full-set API rows (Jev, Sage, wity-1, Fastino GLiNER-2.5-Decide, d1 (Liquid AI)) are not equated. OpenAI Decisions (gpt-6-luna) ran later on the fresh sealed API set A5 ∪ P (600 items; A4 had already been sent to that provider) and is equated the same way with A5's own offsets (+5.78 Intelligence, +6.15 Calibration, same pool).
  • A sealed item is scored in at most three releases, then retired. An item used in one release is not drawn again in the next. If coverage minimums cannot be met, the release waits for newly reviewed items.
  • The selection seed is committed (SHA-256) before any inference, and the draw is a deterministic function of policy, seed and item id.

API-exposure rule

  • An external exposure is any sealed item sent to an endpoint we do not control: closed APIs, and open-weights models reached through third-party hosts, routers or a submitter's endpoint. Timeouts count as exposure.
  • From v1.6.1 hosted models receive the full item set (S plus P); items they had already answered were reused and only the missing sealed items were sent. Hosted models are re-measured at most once every three refresh releases unless a verified new model version ships. Between measurements they keep their last score with its measurement date.
  • Every externally sent sealed item is logged before dispatch. An item exposed to a provider is never scored again for that provider; once two different providers have received it, it retires for everyone. In v1.6.0 Jev 1.13.0 and Fastino GLiNER-2.5-Decide both received the same A, so A retires globally. With the v1.6.1 amendment every v1.6.0 sealed item has been seen by at least two providers, so the whole v1.6.0 sealed draw is retired for future releases.

Public-versus-sealed gap penalty

Intelligence is half public, half sealed. A system whose public score exceeds its sealed score by more than the field-median gap (G_med = 2.6 points) plus 8 points loses one Intelligence point per excess point. Hosted APIs are compared on P versus A against the same self-hosted systems' P-versus-A gap (-1.2 points).

Hosted APIs (not equated since v1.6.1)

Hosted APIs that answered the full set are scored exactly like self-hosted systems; no equating offset is applied to them. Category and language values stay raw. Exception on the live boards: API rows re-run on the fresh API set A4 (v1.7.7) answered A4 ∪ P (600 items) and are equated with the A2/A3 supplement method; OpenAI Decisions the same way on A5 ∪ P (v1.7.10; see Open weights and API offerings above).

Headline and Composite

Capability = mean(Intelligence, Calibration) for systems within twice the Jev 1.13.0 cost and median latency (the official caps; the sliders change only your view). Composite (option A): Weighted harmonic mean of Intelligence, Calibration, Speed and Cost, multiplied by squared penalties when Intelligence is below its configured floor, or Speed or Cost is below 50. Equal 25% weights do not mean an arithmetic average. It leads the API board and is secondary on open weights. Request types Choice, Noul and Score weigh equally; tiers weigh easy 0.1, standard 0.2, hard 0.4, judge 0.3.

Latency. API models are timed end to end from our measuring host, network included. Self-hosted models are timed on our own hardware over loopback (on-GPU, no network), so their measured value is not directly comparable with API latency. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms (est. incl. network (2x + 150 ms)). Tables show the measured value first, tagged “on-GPU, no network”, and the estimate second. A model author can provide a hosted endpoint; we then time it exactly like an API, network included, with no assumption (on request: info@productivity-boost.com, or note it in your submission). From JevBench v1.7.0 (method v1.7), a decision that takes longer than 10 seconds (measured time, not the estimate) counts as a wrong answer and as zero calibration for that item; tables show the measured p95 and how many decisions were over 10 s.

Jev-class caps use a fixed reference: Jev 1.13.0 as measured in v1.5 (p50 0.62 s, USD 0.0323 per 1,000 decisions); caps = 2x (1.23 s, USD 0.0646). Jev's own v1.6 p50 is 0.24 s.

Costs of v1.6-measured systems carry each system's published v1.5.4 cost per 1,000 decisions (pricing rules unchanged; v1.6 item lengths differ) unless the row says otherwise; Fastino's is an estimate from its published tariff and measured tokens.

Noul decisiveness and Score baseline (addendum B)

Scored with method option B (scorer setting O1S), selected on 3 Oct 2026 after the v1.6 scores were known and disclosed as a post-results change. Each split × type competence is clipped at 0 before the type weighting, so a type answered no better than chance counts as chance instead of negative. The Score chance baseline is the error of always predicting the mid-scale level, so a flat know-nothing distribution earns about 0. A Score cell whose golds all sit at mid-scale keeps the v1.5 random-level baseline; this only occurs in small breakdown and bootstrap cells. Calibration is unchanged. This run: G_med = 2.57 points.

Failed requests and very long items

23 items of this draw are very long (about 77,000 to 81,000 input tokens). Systems with a shorter context window refuse them; Jev-Omni runs out of GPU memory on them on its listed RTX 6000 (48 GB) recipe; decider-4b v2's server truncates them to about 32,800 tokens and answers; Plumb-4B reads them in full. Each outcome is the system's own recipe and is scored as such. A failed, refused or unparseable answer counts as wrong for Intelligence and stays in the denominator; it does not enter Calibration (the v1.5 rule, applied to every system). Failed answers per system: AJev lm6q 15/1500 · Aleria Judge 1.1.1 15/1500 · allec-oavg 5/1125 · Amazon Strands Decider 26B-A4B 15/1500 · Amber 4B 1/1125 · Aplomb 1 5/1500 · APUS-OpenJev-v1-35B-A3B 23/1500 · APUS-OpenJev-v1-9B 23/1500 · Arcbright Elf 1 12B 15/1500 · Arcbright Elf 1 12B (r2) 0/1500 · Arcbright Elf 1 31B 15/1500 · ArseneLupin v1.1 15/1500 · AutoJev-27B 23/1500 · basal-omni-1.0-ultra 15/1500 · Bespoke Nimble 9B v3 23/1500 · Bev / Bonsai 27B 23/1500 · BAAI bge-reranker-v2-m3 0/1500 · blink-4b 2/1500 · Blink v0.3 26B-A4B 29/1500 · Bobcat 1.3 15/1500 · Bobcat Flash 1.2 13/1500 · Bongard-mini 15/1500 · Certo v1 0/1500 · Jev-Style-2B 5/1500 · Chon-1.0-12B 540/1500 · Chon-1.0-12B-Fast 0/1500 · Chon-1.0-E4B 0/1500 · ClassOne Gemma 4 E2B 23/1500 · ClassOne Gemma 4 E4B 29/1500 · ClassOne Qwen 3.5 2B 29/1500 · ClassOne Qwen 3.5 4B 29/1500 · ClassOne Qwen 3.5 9B 23/1500 · Clef 0/1500 · Clef-Flash 0/1500 · Clef-omni 0/1500 · CLM-8B 1/1500 · CoCo-Decision-4B 23/1500 · Compass 0.2.0 0/1500 · cribl-decision-1.0-27B 5/1500 · cribl-decision-1.0-9B 0/1500 · cribl-decision-2 5/1500 · cribl-decision-2-mini 5/1500 · Cygnet 23/1500 · D1A on Gemma 4 E2B 15/1500 · D1A on Gemma 4 E4B 15/1500 · decider-12b v2 23/1500 · decider-12b v1, stock Gemma-4-12B-it 23/1500 · decider-2b 0/1500 · decider-35b-a3b 0/1500 · decider-4b v2 0/1500 · decider chat on Gemma-4-31B-it 23/1500 · Decisio 12B tuned commit 15/1500 · Decisio 12B tuned commit-think 0/1500 · decisio v0.8.0 on gemma-4-12B-it 23/1500 · decisio v0.9.0 on gemma-4-12B-it 23/1500 · decisio v0.8.0 on gemma-4-31B-it 23/1500 · Decisio v0.9.0 on Qwen3.6-35B-A3B 15/1500 · Decision 2.0 Vega 27B 23/1500 · Decision 2B 23/1500 · Decision 4B v1.1 23/1500 · Decision 4B v1.2 23/1500 · Decision Fast 23/1500 · Decision-King v1.1 6/1500 · Decision 2.0 Lux 9B 15/1500 · DecisionNode Edge 0.9.2 15/1500 · DecisionNode Edge Mini 0.9.2 15/1500 · decisor-4b 0/1500 · deck-4B v1.0 23/1500 · deck-31B 23/1500 · Deem 0.8B v1 23/1500 · Diffusion Jev 26/1500 · DOPEST 1.0 31B 0/1500 · Eikos-27B 23/1500 · baselabs elephant-1 15/1500 · Decision-4B 23/1500 · EXAONE-4.0-1.2B-JEV v0.3 23/1500 · Fastino GLiDE 1.2 15/1497 · GEV-26B-Decide 23/1500 · Arbiter-4B 15/1500 · Granite Switch 4.1 30B 5/1500 · Granite Switch 4.1 8B 5/1500 · Alibaba GTE Reranker ModernBERT-base 0/1500 · Gutsy 0.8B v0.3 29/1500 · H2O-Lightning-31B v1.0.7 15/1500 · H2O-Lightning-4B v1.1 0/1500 · H2O-Lightning-4B v1.4.1 0/1500 · Hopper 22/1500 · Hopper 12B trained 23/1500 · Imajev-2B 15/1500 · Imajev-4B 23/1500 · Imajev-9B 15/1500 · iota-decide-4b 0/1500 · iota-decide-4b-v0.3 0/1500 · iota-decide-4b-v0.4 0/1500 · JADE 392/1500 · janus 4B 23/1500 · JEB-35B-A3B 15/1500 · Jebadiah 27B 24/1500 · Jebadiah 9B v2 15/1500 · Jebadiah 9B v2.1 15/1500 · Jeetu Decider 2B 15/1500 · Jeff-1.0-Large 21/1500 · Jeff-1.1-Large 0/1500 · Jeff-1.1-Medium 0/1500 · Jeff-1.1-Qwen-4B 0/1500 · Jeffan-31b-1710 15/1500 · Jeffan-4B-2010 15/1500 · Jev 1.13.0 23/1500 · Jev 1.14.0 15/1500 · JEV-27B 23/1500 · jev-local 23/1500 · Jev-Omni 23/1500 · JEV Qwen3.5-9B Base NVFP4 23/1500 · JevK5 v0.3 23/1500 · JevK5 v0.2.0 23/1500 · JevOne 0/1500 · jevora-9b-v1 49/1500 · Jez-large-1.0.0 15/1500 · Jez-small-1.0.0 15/1500 · jiwo 0.8b 29/1500 · jiwo 4b 29/1500 · baselabs joey-1 5/1500 · JPT-35B-A3B 23/1500 · JPT-4B 15/1500 · jqv 23/1500 · jul fast 6/1500 · Julia-1 17/1500 · Kahn1 4B 23/1500 · kev 0.5B 32/1500 · kev 0.6B 30/1500 · Kev 27B 23/1500 · kev 4B 29/1500 · kev 8B 25/1500 · Kev-9B v2 15/1500 · GLiClass v4 9B 15/1500 · KnowLine 26B-A4B-Gen2.5 15/1500 · KnowLine 31B-Gen1 15/1500 · KnowLine 35B-A3B-Gen1.5 0/1500 · KnowLine 4B-Gen5.5 0/1500 · Krino-2B 0/1500 · Laya 0/1500 · Laya multilingual 0/1500 · Laya typed-decisions 0/1500 · lev 25/1500 · lev-350m 23/1500 · Liquid AI d1-3B 0/1500 · Liquid AI d1-omni-600M 0/1500 · LitJev 23/1500 · local-jev Qwen3.5-4B 0/1500 · FrontiersMind Lumma-fev-27B 15/1500 · Lychee-27B 15/1500 · Malkuth-2B 23/1500 · Malkuth-4B 23/1500 · baselabs mammoth-1 15/1500 · Manchego v2.1 23/1500 · Messier One v0.2 0/1500 · Messier One v0.4 0/1500 · MetaDecision-31B-Rain 0/1500 · metask-jev-4b 23/1500 · Metask rain 4B 0/1500 · metask-jev-rain-12B 0/1500 · metask-jev-rain-12B-trained 12/1500 · metask-jev-rain-31B-trained 0/1500 · metask-jev-rain-31B-trained service v3 0/1500 · Qwen3.5-0.8B Decision Model 0/1500 · Mirror 592/1500 · modernbert-jev v2 0/1500 · Mixedbread mxbai-rerank-base-v2 0/1500 · Nagi-ENORMOUS 27B 15/1500 · Nagi-HUGE 12B 15/1500 · Nemotron Deciffusion 8B v2 0/1500 · Nemotron Diffusion 8B 23/1500 · Nev-2B 15/1500 · Bespoke Nimble 9B 23/1500 · Nimble-v4-max 16/1500 · NInfer Qwen3.8-27B NVFP4 23/1500 · NInfer Qwen3.8-Flash-Next mixed 23/1500 · Noma 0/1500 · oenogarum 5/1125 · open-alternative-jev 0/1500 · open-jev-deberta-v3-large 73/1500 · Open Jev JSON Canvas 23/1500 · Open-Jev 27B v1.1 23/1500 · Open-Jev 2B 23/1500 · Open-Jev 9B 23/1500 · OpenDecision 0/1500 · OpenJev 24/1500 · openJev Verdict 0/1500 · openJev Verdict 1.4 0/1500 · OpenSourceJev 23/1500 · Philly Decider v2 2/1500 · Phocinae-Largha-150M-v1 375/1500 · Piccolo 27B 5/1500 · Piccolo 4B 5/1500 · Piccolo 9B 5/1500 · pjev-2b 15/1500 · Plumb-4B 2/1500 · pplx-decider-flash-v1.2 15/1500 · Perplexity Decider v1.1 27B 24/1500 · pplx-decider-v1.2 15/1500 · PurpleMIST-Flash-1.0 15/1500 · PurpleMIST-Mini-1.0 15/1500 · Quyet-1.0-Large 0/1500 · Quyet-1.0-Medium 0/1500 · Quyet-1.0-Small 0/1500 · Quyet-1.0-Small-EN 0/1500 · Quyet-1.0-Tiny 0/1500 · Qwen3-Reranker-4B 0/1500 · Qwen3.5-9B Jev-like data-mix v2 23/1500 · Raw Phi-4 mini direct logits 23/1500 · Raw Qwen3 0.6B direct logits 23/1500 · Raw Qwen3 1.7B direct logits 23/1500 · Raw Qwen3 4B Instruct 2507 direct logits 23/1500 · Raw Qwen3 8B direct logits 23/1500 · reflex 4B 2/1500 · Reflex Instinct 0.6B 0/1500 · René-1 31B FP8 23/1500 · Respan Span-02 15/1500 · Decider-4B 15/1500 · RSI-Jev v6.1-VL 27B 29/1500 · RYOTIDE-Gemma 15/1500 · RYOTIDE-Qwen 2/1500 · RYOTIDE-Qwen9 21/1500 · Saif 4B 0/1500 · Seb-9B 15/1500 · Seb-9B-v2 15/1500 · SemIf, formerly OpenJev 23/1500 · shisa-de-1 16/1500 · Sieve-31B-v2 15/1500 · SimpleJev 37/1500 · SimpleJev Qwen3.8-27B 23/1500 · smalljev semantic-v9 0/1500 · Sors-2B 5/1500 · spark-s1-4b-v6 0/1500 · SPX-CD-Omni 23/1500 · Standard One 3B SH 15/1500 · Standard One 3B SH v4 15/1500 · Standard One 8B SH 29/1500 · Standard One 8B 23/1500 · StartLux-Decision v1.1 27B AutoThink 27/1500 · StartLux-Decision v1.1 27B Instant 15/1500 · Surogate Rune 26B-A4B v3 23/1500 · swanOne 23/1500 · system-one 23/1500 · Tacet Sonata 3/1500 · TDE general large v0.1 0/1500 · Tern-35B-A3B 15/1500 · torchcast-decision-12b 23/1500 · torchcast-decision-27b 23/1500 · Torchcast Decision 31B v1.0 0/1500 · Torchcast Decision 4B (v1.1) 0/1500 · Torchcast Decision 4B (v1.2) 0/1500 · Torchcast Decision 4B (v1.3) 0/1500 · Turion Jev 31B 15/1500 · typecastlm 0/1500 · TypeCastLM 1.4.0 0/1500 · Umpire 35B-A3B 5/1500 · Uniphore-4B-Decider 15/1500 · Vega-U (Qwen3.5-0.8B reader) 15/1500 · Vega-U (Qwen3.5-4B reader) 15/1500 · Vega 0.8B 0/1500 · Vega 4B 29/1500 · Vera Large 15/1500 · verdict-small 0/1500 · Verdin-4B 15/1500 · VisionJEV 1.0.0 5/1500 · vLLM-SR d3 15/1500 · vLLM-SR d3-mini 15/1500 · vLLM-SR d3-turbo 15/1500 · Wald 26B-A4B 0/1500 · Wald 4B 8/1500 · Wald 4B v2.1 0/1500 · WaterSheep 1/1500 · watt-flash-0.1 23/1500 · WeMM-JEV 5/1500 · Winnow-12B Q8 23/1500 · Xevn-4B 15/1500 · Xor 26B-A4B 23/1500 · ZeroEntropy zerank-2 0/1500 · z3-mini-soup-preview 5/1500.

Calibration basis

Systems that return a full probability distribution are calibrated on all components (top-label error, plus distribution distance for Choice and ranked-probability error for Score).

Overnight full re-measure (4–5 Oct 2026)

Every system with a reproducible recipe was re-run on the v1.6.0 pool overnight with the same pinned inputs and scorer (method option B / O1S). This page uses scoring round score-v161-3 (v1.6.1) (2026-10-06 00:39:36 UTC). Only complete runs (1,500 items self-hosted, the full API input for hosted APIs) are ranked; partial runs are never ranked, and systems not yet re-measured keep their dated v1.5.x score in the separate table.

  • Scores are the official v1.6.0 scorer (score_v16.py, method option B / O1S, bootstrap B = 1,000) over complete outputs only; each scoring round is kept separately.
  • Hosted APIs are scored on the same full item set as self-hosted systems (S 1,200 + P 300, v1.6.1) and are not equated.
  • GPU-class deviations: reproducible recipes used the hardware listed per row, including H100 for large fast-lane decoders and RTX 6000 for the baseline; hardware differences remain in measured latency. The standard x2 + 0.15 s self-hosted adjustment is an assumption, not a hardware normalization.
  • Wity auto remains outside the latency cap and stays ranked in Composite A. OFF and ALWAYS are unranked variants of the AUTO main row.
  • 6 further measured candidates await a separate publication decision.
  • Sage 1.3.0 (Levanto Labs) was measured on 5 Oct from Sandy (Helsinki), text only. Cost uses the Levanto list tariff (USD 0.05/M input, USD 10/M output; levanto.ai/pricing, read 5 Oct) over all 600 answered rows: USD 0.0766/1,000 decisions. Superseded by the v1.6.1 cost rule (common item set): USD 0.0247/1,000 decisions, inside the Jev-class cost cap.
  • A fresh Monday Jev 1.13.0 re-check on A3 measured p50 0.239 s and equated Capability 77.5 versus the official 76.5, within the confidence interval; the official v1.6.0 Jev row is retained.

Display note on the overnight text above (frozen release data). Exception since v1.7.7: the 17 API rows re-run on A4 ∪ P (600 items) are equated (+2.57 Intelligence, +3.35 Calibration); see “Full API re-run (A4, v1.7.7)” under Open weights and API offerings. Full-set API rows (Jev, Sage, wity-1, Fastino GLiNER-2.5-Decide, d1 (Liquid AI)) are not equated. OpenAI Decisions (gpt-6-luna) ran later on the fresh sealed API set A5 ∪ P (600 items; A4 had already been sent to that provider) and is equated the same way with A5's own offsets (+5.78 Intelligence, +6.15 Calibration, same pool).

Supplementary API draws A2 and A3

History: hosted APIs were first measured on API subsets (A, then A2 and A3) and equated to the self-hosted scale in v1.6.0. From v1.6.1 they answer the same full set as self-hosted systems, so no equating is applied; the earlier subsets A, A2 and A3 are retired.

Per-model exposure counts (hosted and author-hosted endpoints)

System (provider)v1.6 sealed items sentScored sealed setStatus
jev-1.13.0 (typesafe)1200S (1,200 sealed) + P (300 public)full set since v1.6.1 (5 Oct 2026); earlier API-subset answers reused, only missing sealed items sent; whole v1.6.0 sealed draw retired

Provenance

Measured on the v1.6 pool (run completion day, UTC): 2026-10-02: AutoJev-27B, BAAI bge-reranker-v2-m3, Certo v1, CLM-8B, Cygnet, decider-2b, decider-35b-a3b, decider-4b v2, Decision 4B v1.1, Decision 4B v1.2, Eikos-27B, Alibaba GTE Reranker ModernBERT-base, Hopper, Imajev-4B, jev-local, Jev-Omni, JEV Qwen3.5-9B Base NVFP4, JevK5 v0.3, JevK5 v0.2.0, JevOne, jqv, kev 0.5B, kev 0.6B, kev 4B, kev 8B, Laya, LitJev, local-jev Qwen3.5-4B, Malkuth-2B, Malkuth-4B, metask-jev-4b, Qwen3.5-0.8B Decision Model, Mirror, Mixedbread mxbai-rerank-base-v2, Bespoke Nimble 9B, NInfer Qwen3.8-27B NVFP4, NInfer Qwen3.8-Flash-Next mixed, open-alternative-jev, open-jev-deberta-v3-large, Open Jev JSON Canvas, Open-Jev 2B, Open-Jev 9B, OpenDecision, OpenJev, openJev Verdict, openJev Verdict 1.4, OpenSourceJev, Plumb-4B, Qwen3-Reranker-4B, Qwen3.5-9B Jev-like data-mix v2, Raw Phi-4 mini direct logits, Raw Qwen3 0.6B direct logits, Raw Qwen3 1.7B direct logits, Raw Qwen3 4B Instruct 2507 direct logits, Raw Qwen3 8B direct logits, reflex 4B, SemIf, formerly OpenJev, SimpleJev, smalljev semantic-v9, Standard One 8B, swanOne, system-one, typecastlm, verdict-small, Winnow-12B Q8, ZeroEntropy zerank-2 · 2026-10-04: Bev / Bonsai 27B, Clef, Clef-Flash, decider-12b v2, decider-12b v1, stock Gemma-4-12B-it, Deem 0.8B v1, Decision-4B, H2O-Lightning-4B v1.1, janus 4B, Laya multilingual, Laya typed-decisions, lev, Manchego v2.1, Nemotron Diffusion 8B, Open-Jev 27B v1.1, Quyet-1.0-Large, Quyet-1.0-Medium, Quyet-1.0-Small, Quyet-1.0-Small-EN, Quyet-1.0-Tiny, torchcast-decision-12b · 2026-10-05: Decision 2B, Decision Fast, deck-4B v1.0, deck-31B, Jev 1.13.0, lev-350m, spark-s1-4b-v6 · 2026-10-06: Aplomb 1, Bespoke Nimble 9B v3, decider chat on Gemma-4-31B-it, decisio v0.8.0 on gemma-4-12B-it, decisio v0.8.0 on gemma-4-31B-it, Decision 2.0 Vega 27B, Diffusion Jev, GEV-26B-Decide, Hopper 12B trained, JADE, Jebadiah 27B, JEV-27B, JPT-35B-A3B, Kev 27B, Perplexity Decider v1.1 27B, SimpleJev Qwen3.8-27B, SPX-CD-Omni, Surogate Rune 26B-A4B v3, torchcast-decision-27b, Xor 26B-A4B · 2026-10-07: APUS-OpenJev-v1-35B-A3B, APUS-OpenJev-v1-9B, Bobcat Flash 1.2, ClassOne Gemma 4 E2B, ClassOne Qwen 3.5 9B, CoCo-Decision-4B, decisio v0.9.0 on gemma-4-12B-it, EXAONE-4.0-1.2B-JEV v0.3, jul fast, Kahn1 4B, Messier One v0.2, René-1 31B FP8, Tacet Sonata, TypeCastLM 1.4.0, WaterSheep, watt-flash-0.1 · 2026-10-09: decisor-4b, Jeff-1.0-Large, Metask rain 4B, metask-jev-rain-12B, Wald 4B · 2026-10-10: Blink v0.3 26B-A4B, ClassOne Gemma 4 E4B, ClassOne Qwen 3.5 2B, ClassOne Qwen 3.5 4B, Clef-omni, Gutsy 0.8B v0.3, jiwo 0.8b, jiwo 4b, Liquid AI d1-3B, Liquid AI d1-omni-600M, modernbert-jev v2, RSI-Jev v6.1-VL 27B, RYOTIDE-Qwen9, Standard One 8B SH, Vega 0.8B, Vega 4B, Wald 4B v2.1 · 2026-10-12: Krino-2B, metask-jev-rain-12B-trained, metask-jev-rain-31B-trained, Standard One 3B SH, Torchcast Decision 4B (v1.1), Uniphore-4B-Decider, vLLM-SR d3 · 2026-10-13: Julia-1, Lychee-27B, metask-jev-rain-31B-trained service v3, Phocinae-Largha-150M-v1, Reflex Instinct 0.6B · 2026-10-14: Messier One v0.4, Respan Span-02 · 2026-10-15: Nemotron Deciffusion 8B v2, Nev-2B · 2026-10-16: ArseneLupin v1.1, blink-4b, Bongard-mini, Compass 0.2.0, Decision 2.0 Lux 9B, Imajev-2B, Imajev-9B, JEB-35B-A3B, Jebadiah 9B v2, Jebadiah 9B v2.1, Jev 1.14.0, jevora-9b-v1, JPT-4B, Kev-9B v2, Nagi-ENORMOUS 27B, Nagi-HUGE 12B, Noma, RYOTIDE-Gemma, RYOTIDE-Qwen, shisa-de-1, StartLux-Decision v1.1 27B AutoThink, StartLux-Decision v1.1 27B Instant, TDE general large v0.1 · 2026-10-17: basal-omni-1.0-ultra, PurpleMIST-Flash-1.0 · 2026-10-18: Torchcast Decision 4B (v1.2), Vega-U (Qwen3.5-0.8B reader) · 2026-10-19: Chon-1.0-12B, pplx-decider-flash-v1.2, vLLM-SR d3-turbo · 2026-10-20: Arcbright Elf 1 12B, pplx-decider-v1.2, Seb-9B · 2026-10-21: Arbiter-4B, iota-decide-4b, Nimble-v4-max, Decider-4B · 2026-10-22: AJev lm6q, Aleria Judge 1.1.1, Arcbright Elf 1 31B, baselabs elephant-1, H2O-Lightning-31B v1.0.7, Jez-large-1.0.0, Jez-small-1.0.0, FrontiersMind Lumma-fev-27B, baselabs mammoth-1, pjev-2b, PurpleMIST-Mini-1.0, Tern-35B-A3B, Verdin-4B · 2026-10-23: Bobcat 1.3, Fastino GLiDE 1.2, KnowLine 26B-A4B-Gen2.5, KnowLine 35B-A3B-Gen1.5, KnowLine 4B-Gen5.5, Seb-9B-v2, Standard One 3B SH v4, Torchcast Decision 4B (v1.3), Vega-U (Qwen3.5-4B reader), Wald 26B-A4B · 2026-10-24: Amazon Strands Decider 26B-A4B, DecisionNode Edge 0.9.2, DecisionNode Edge Mini 0.9.2, Jeffan-31b-1710, Jeffan-4B-2010, KnowLine 31B-Gen1, Turion Jev 31B, Vera Large, vLLM-SR d3-mini, Xevn-4B · 2026-10-25: Chon-1.0-12B-Fast, Chon-1.0-E4B, Decisio 12B tuned commit, DOPEST 1.0 31B, Jeetu Decider 2B, Sieve-31B-v2 · 2026-10-26: D1A on Gemma 4 E2B, D1A on Gemma 4 E4B, Decisio v0.9.0 on Qwen3.6-35B-A3B, iota-decide-4b-v0.3, Jeff-1.1-Medium, GLiClass v4 9B, Philly Decider v2 · 2026-10-27: Arcbright Elf 1 12B (r2), Decisio 12B tuned commit-think, H2O-Lightning-4B v1.4.1, Jeff-1.1-Large, Jeff-1.1-Qwen-4B, MetaDecision-31B-Rain, Torchcast Decision 31B v1.0 · 2026-10-28: Jev-Style-2B · 2026-10-29: cribl-decision-2, cribl-decision-2-mini, Decision-King v1.1, iota-decide-4b-v0.4, baselabs joey-1, Saif 4B, Umpire 35B-A3B, VisionJEV 1.0.0, WeMM-JEV · 2026-10-30: allec-oavg, Amber 4B, cribl-decision-1.0-9B, oenogarum, Piccolo 27B, Piccolo 4B, Piccolo 9B, Sors-2B · 2026-10-31: cribl-decision-1.0-27B, Granite Switch 4.1 30B, Granite Switch 4.1 8B, z3-mini-soup-preview.

Aggregate files: results sha256 5cd8c1332226ea9c179883d2285c0d8b8cc593ae38f2645197364a7ddd9c83c1 · categories sha256 0159c4828b3e8620148af06848ef0a5ccfc7d7f0c11aec02460d844dc0c62689 · dated carry sha256 334e851944e92383ab3071e52b3012040c7023d0fe236395953169ebee8f0459 · live category union (data/raw/benchmarks/jevbench/v1.6/jevbench-v1.6.1-category-cells.json) sha256 99aa71d219028acaa5c306cf3a087c9d3d4d844667ef9b76b61f6063383d1edb. Scoring source sha256 e15347783b8bedeeddd9017a831608a957f5bf0386f05902c50161ce0468f911. The method, release data and carry artifact are independently hashable.

Release history

The board above is always the current merged board: every system with its latest measurement. Each release stays readable as published.

Since v1.7.0: a decision that takes longer than 10 seconds (measured time) counts as a wrong answer and as zero calibration for that item, and Speed uses all decisions. The v1.6 values stay visible on every row.

Historical v1.3.0 board: weightings, per-task grid, topic radars and held-out diagnostics

Open this section to load the earlier public-only board and diagnostics.

Frequently asked questions

What is JevBench?
JevBench is Benchmark Heaven’s benchmark for AI decision models, including Jev-compatible systems: state and a bounded rubric in, a typed answer out. It measures Intelligence, Calibration, Speed and Cost.
Which model ranks #1 on JevBench?
baselabs mammoth-1 (27B) ranks #1 on the open-weights board by Capability Score (85.2) among 217 ranked open-weights models in v1.8.7. API offerings are ranked separately on the API leaderboard (https://benchmarkheaven.com/jev-models/api). Composite is secondary.
What is the best open-weight Jev alternative?
Perplexity Decider v1.1 27B (Qwen3.8-27B, noncausal decision readout) is the highest eligible open-weight Jev-compatible alternative in v1.8.7, with Capability 82.7 and rank #5 on the open-weights board. Review its licence and source before deployment.
How is the Capability Score computed?
Capability is the arithmetic mean of Intelligence and Calibration. Official ranks include only ranked models within the cost cap (0.06459465517241379 USD per 1,000 decisions) and median-latency cap (1.2329566404223442 seconds). The reference is Jev 1.13.0. Composite is secondary.
Why does JevBench have a 10-second limit per decision?
A decision model runs inside live software, where an answer that arrives half a minute late is usually as useless as no answer. A median time hides this: a system can answer most items in a fraction of a second and still spend 30 seconds or more on many of them. From v1.7.0 the limit makes the slow tail count, in Intelligence (a late decision counts as failed) and in Speed (all times count).