cribl-decision-1.0-9B (Cribl)v1.7 · 10 s rule

Composite (secondary)

52.5 · Not ranked on this board

Cost and measurement conditions

$0.0068 per 1,000 decisions (estimated)

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; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 457,748 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00679680125958427/1,000; old list-price estimate USD 0.0307419744795164/1,000 (qwen/qwen3.5-9b at USD 0.1/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00679680125958427/1,000. Estimated, not a bill. USD 0.00679680126 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0125854348 over all 1,500 items).

p50 latency: 0.2 s

Self-hosted latency estimate under the shared method (assumption, not measured)

Measured 30 Oct 2026 on an evaluator-owned Verda VM in FIN-03 (Finland, EU), one NVIDIA RTX PRO 6000 Blackwell Server Edition, nominal 96 GB, image 24.04.cuda13.0. FIRST-PARTY IMPLEMENTATION OF THE AUTHOR'S DOCUMENTED READOUT, with declared extensions: cribl-ai/cribl-decision-1.0-9B adapter at 73abdfbecf0619fdc02bd2d769460b353d3de30a, attached to Qwen/Qwen3.5-9B at c202236235762e1c871ad0ccb60c8ee5ba337b9a. This is a PEFT adapter plus base model, not a standalone full fine-tuned checkpoint. The author supplies a single-question demo with ten implemented labels and describes a larger label table. The shim preserves its Context/Question/Options/Answer prompt, native state serialization, original option order, last-token hidden state, candidate lm_head projection, float logits and candidate-restricted softmax. No chat template, prompt sanitization, truncation or generated text. It loads explicit Qwen3_5ForConditionalGeneration plus PEFT, then get_base_model().model.language_model and lm_head, with BF16 CUDA base and ordinary PEFT adapter dtype behavior. Stock efficient SDPA is selected by the locked runtime; no attention override, cache override, quantization, CPU fallback, host retry or answer repair. Multiple questions use separate serial passes; no invented multi-slot packing. DECLARED EXTENSIONS: noul maps proposition/negation in true/false order, with instructions and literal Not (instructions) as absent-description fallback; score preserves ordered levels and returns the complete distribution keyed by zero-based index strings. The card's 255-label single-token constraint is interpreted by A-Z followed by the first 229 lexically ordered uppercase pairs that tokenize singly, ending at JT; split-token pairs are skipped. This is a disclosed interpretation, not proof of an unavailable complete author training table. Public pod parity compares token IDs and bit-exact probabilities with extracted native prompt/readout functions. 8,192 is training length, not a refusal cap; the demo forwards every unpadded token without a native total-length check. Architectural 262,144 positions do not establish empirical capacity. No context cap is added; genuine unsupported type/candidate contracts may return HTTP 422, while framework/OOM faults remain HTTP 500. First-party HTTP input usage sums every actually forwarded prompt, including repeated state across question passes; output_tokens is 0. The author supplies no HTTP server or usage billing contract. The frozen TypeSafe adapter and driver retain the 120 s transport timeout. Locked runtime: torch 2.11.0, CUDA 13.0, transformers 5.19.0, tokenizers 0.23.3, PEFT 0.21.2, safetensors 0.8.0; source/runtime/tokenizer hashes retained. Source review found no benchmark-item lookup in the selected author readout; upstream framework internals were not exhaustively audited. Numerical and author-code execution occurred only on the disposable rented pod, with verified local base/adapter, read-only mounts, empty credentials, offline flags and network cut. Actual-model public smoke passed with native/offline usage equality on its public fixtures, separate warm-ups and fresh timed bodies from 256 through 80,554 rendered tokens. All documented types and widths through 255, Unicode JSON, special-token and multi-question fixtures passed with native prompt/readout parity. No native length-refusal probe is claimed. Public equality does not erase the 56 scored usage mismatches. These are synthetic observations, not scored latency aggregates. Separate public pre-warm passed. The complete 1,500-request run took 126.39241480827332 s; raw output sha256 6ccacaa3976d98611717ac669ab06710d0708d20fcdad889d3889836c9f88193 was verified. Post-run health was ready. The VM was closed and temporary node and local copies deleted, with no cleanup errors reported. GPU rental spend is separate from the benchmark's model-cost estimate.