Quyet-1.2-Smallv1.7 · 10 s rule

Data: JevBench v1.8.15, published 2026-11-17 · method v1.7 (10 s rule), measured under the rule

Capability 56.6 · Not ranked in the current release

Added after the v1.8.10 ranking review; it is ranked at the next review.

Last measured: 2026-11-16 · measurement release: v1.8.14

Apache-2.0

Base model: google/gemma-4-E4B-itsource

intelligence

31.8

calibration

81.4

speed

93.4

cost

88.9

Composite (secondary)

24.6 · Not ranked on this board

Cost and measurement conditions

$0.0024 per 1,000 decisions (estimated)

GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, USD 1.5235/hour; average load 70%. Architecture: small gemma, 4B active parameters; throughput 80585.75 tokens/s (architecture-class calibration, efficient-engine prefill). GPU-time cost USD 0.00235011659878878/1,000; old list-price estimate USD 0.00626521155137676/1,000 (deepinfra:google/gemma-4-E4B-it at USD 0.02/M input and USD 0.1/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00235011659878878/1,000. Estimated, not a bill.

Measured p50 latency: 0.024 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: 0.20 s (est. incl. network (2x + 150 ms))

Self-hosted model: measured p50 0.024 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 = 0.20 s. Model authors can ask us to time a hosted endpoint instead (on request).

x2 + 0.15 s (assumption, not measured)

evaluator-owned Verda GPU VM (RTXPRO6000), offline read-only container (--network none); the author's quyet 1.2.0 package (github.com/ncchinh/quyet @d41fa079) loaded in process by the fixed pod driver (quyet.load(model_dir, device='cuda').predict); full BF16 fine-tune of Gemma 4 E4B (text decoder) + option head (kind student), SDPA, one forward pass per question, no generation, no prefix cache; inputs longer than the model's 23,434-token input limit, or choice questions with more than 256 options, are refused and scored as wrong; official PyTorch 2.14.0 cu130 image + pinned overlay (transformers 5.17.0, accelerate 1.15.0, Python 3.12)

Measured latency over all decisions: p50 0.024 s · p95 0.041 s

— decisions over 10 s.

A decision that takes longer than 10 seconds (measured time) counts as a wrong answer and as zero calibration for that item: each type's calibration is multiplied by the share of its decisions answered within 10 s. The limit uses the measured time (raw). The displayed adjusted time for self-hosted rows (2 × + 0.15 s) is an assumption and is not used for the limit.

Model and serving disclosure

Paid public evaluation. Measured on the 27 Oct draw on an evaluator-owned Verda GPU VM (RTXPRO6000), offline with the author's serving stack at the pinned commit. Failed or unanswered requests count as wrong answers. Decision model: full fine-tune of Gemma 4 E4B with a separate option head; choice via the option head, score and noul via vocabulary-symbol scores with fixed per-type temperatures; 0 output tokens.

Published source