Sieve-31B-v2v1.7 · 10 s rule
Data: JevBench v1.7.5, published 2026-10-25 · method v1.7 (10 s rule), measured under the rule
Capability 81.1 · rank #7 of 163 on the open-weights board
Last measured: 2026-10-25 · measurement release: v1.7.5
Apache-2.0 per model card and LICENSE on weights and author library; base google/gemma-4-31B-it public and ungated under its own terms. No serving code redistributed here.
intelligence
68.7
calibration
93.4
speed
85.6
cost
50.9
Composite (secondary)
70.7 · Composite rank #9
Cost and measurement conditions
$0.043 per 1,000 decisions (estimated)
ESTIMATE (*est.), base google/gemma-4-31B-it priced through frozen pricing_v15 (aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7) as google/gemma-4-31b-it at USD 0.09/M input and USD 0.34/M output. Native usage: 714,573 input and 0 output tokens across 1,485 answered items, matching the offline count exactly. Under the published zero-output convention for native-probability rows the output term is zero. The 15 HTTP 422 refusals carry no usage. Common 1,485-item basis: USD 0.043307455/1,000 decisions, 0.670449504 times the unchanged USD 0.06459465517241379 cap. All 1,500: USD 0.042874380/1,000, 0.663745009 times cap. Estimated model tariff, not charged rental or a host quote.
Measured p50 latency: 0.138 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: 0.43 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.138 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 = 0.43 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
Evaluator-owned rented Verda GPU, one NVIDIA RTX PRO 6000 Blackwell Server Edition 96 GB, Finland, EU. One attempt, measured 25 Oct 2026. Author engine behind our own loopback shim in an offline, read-only container with empty environment and hash-checked model and code. Complete 1,500-request run; scored loop 237.8 s. About USD 1.31 GPU rental, separate from the estimated model tariff.
Measured latency over 1,500 decisions: p50 0.138 s · p95 0.247 s
0 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.