Blink v0.3 26B-A4B
Data: JevBench v1.6.39, published 2026-10-22
Capability 69.9 · rank #27 of 139 on the open-weights board
Last measured: 2026-10-10 · measurement release: v1.6.7
Apache-2.0 (card tag; no LICENSE file in the repo); base google/gemma-4-26B-A4B-it Apache-2.0
intelligence
49.2
calibration
90.7
speed
91.5
cost
49.6
Composite (secondary)
61.2 · Composite rank #32
Cost and measurement conditions
$0.048 per 1,000 decisions (estimated)
ESTIMATE (*est.), base-model reference. v1.5-M2 base-model reference: Blink v0.3 26B-A4B is a full fine-tune of google/gemma-4-26B-A4B-it quantised to NVFP4 (ModelOpt); the base resolves in pricing_v15 as base_model google/gemma-4-26b-a4b-it -> OpenRouter google/gemma-4-26b-a4b-it, USD 0.09/M input, USD 0.30/M output, from the OpenRouter catalog snapshot 2026-09-25 (jevbench-v15-measure-20260925/receipts/OPENROUTER-MODELS-20260925.json, sha256 f5d0cd8e0f7d2d6514748c99bfcaeacc09775bac7aa7eecf765af4ba04b0dcd5), read by the frozen pricing_v15 on 2026-10-10. Tokens: input = vLLM prompt_tokens, output = 1 per answered question (one decode step, the option letter); both priced. The 29 context-refused decisions carry no tokens and stay in the 1,500-decision denominator.
Measured p50 latency: 0.051 s (on-GPU, no network, loopback; not directly comparable with API latency, which we time end to end from our measuring host, network included)
Ranking and cap use: 0.25 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.051 s on-GPU, no network (timed on our GPU over loopback). Not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and latency cap use the same estimate for every self-hosted row: 2 × measured + 150 ms = 0.25 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
evaluator-owned Lium RTX 5090 32 GB pod (Drobeta-Turnu Severin, Romania), offline after setup (egress cut and proven), plain vLLM 0.30.0 with the model card's own serve command plus ONE serving-only deviation: VLLM_BATCH_INVARIANT=1 (needed to pass the 0.005 repeatability gate; see DETERMINISM.md), served through our loopback shim implementing the card's reference decide() client (T=1.3, softmax over option-letter logprobs); context 32,768 (29 of 1,500 states are 77k-82k tokens and were refused by vLLM = recorded misses)