Sieve-31B-v3v1.7 · 10 s rule

Data: JevBench v1.8.16, published 2026-11-19 · method v1.7 (10 s rule), measured under the rule

Capability 84.9 · Not ranked in the current release

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

Last measured: 2026-11-19 · measurement release: v1.8.16

apache-2.0 (model card; base Gemma terms also apply)

Base model: google/gemma-4-31B-itsource

intelligence

77.4

calibration

92.4

speed

83.1

cost

52.5

Composite (secondary)

72.9 · Not ranked on this board

Cost and measurement conditions

$0.038 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: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, efficient-engine prefill). GPU-time cost USD 0.038445653430234/1,000; old list-price estimate USD 0.0578428005372733/1,000 (google/gemma-4-31b-it at USD 0.09/M input and USD 0.34/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.038445653430234/1,000. Estimated, not a bill.

Measured p50 latency: 0.189 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: 0.53 s (est. incl. network (2x + 150 ms))

Self-hosted model: measured p50 0.189 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 = 0.53 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, offline read-only container (--network none); first-party loopback shim over native LettersEngine at code commit 1e64657c, true package version 0.5.1; three reviewed source substitutions replace metadata deserialization/provenance and restrict the package to letters only. head.pt is never downloaded or unpickled: identical SHA256 metadata is embedded from the prior strict builtins-only decode. BF16 Gemma-4-31B backbone, native full-precision calibrated label softmax; reads=1, reread_max=3000, orders=2, no context truncation/default limit. Actual backbone token concordance covers both option orders, usage sums unpadded passes, output_tokens=0. Native unsupported shapes return 422; GPU OOM is operational failure, not a length refusal. Official PyTorch 2.9.1 CUDA12.8 image plus reviewed offline Sieve supplement.

Measured latency over all decisions: p50 0.189 s · p95 0.394 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.

Breakdowns (not part of the score)

Chance-corrected competence (0 = chance, 100 = perfect) by category on open+sealed. Raw values, not equated; the same cells feed the compare view.

Capability by subject topic

subject topics: not labelled for this draw yet (top-up pending)

Use cases (TypeSafe categories)

Model routing85.2n=148
Legal & compliance89.3n=56
Customer support85.7n=66
E-commerce79.8n=62
Insurance claims84.6n=62
Risk assessment75.7n=48
Financial crime77.8n=63
Feature extraction82.5n=72
Lead generation93.5n=75
Recruiting88.2n=65
Knowledge graphs95.2n=67
LLM guardrails98.0n=71
Moderation92.9n=59
Code linting96.4n=69
Search & retrieval95.0n=67
Science71.2n=42
Advertising82.8n=48
Gaming81.4n=58
Demand forecasting16.1n=66

Cells need at least 30 answered items; † fewer than 30 (indicative only; drawn as hollow, lighter points on the compare radars).

Model and serving disclosure

Paid public evaluation. Measured on the 27 Oct draw on an evaluator-owned rented GPU, offline with the reviewed Sieve runtime. Failed or unanswered requests count as wrong answers. The native LettersEngine evaluates two option orders and reports the summed unpadded input usage, with zero output generation.

Published source