Jeff-1.1-Medium (jgeuter)v1.7 · 10 s rule
Data: JevBench v1.7.8, published 2026-10-26 · method v1.7 (10 s rule), measured under the rule
Capability 81.9 · rank #4 of 167 on the open-weights board
Last measured: 2026-10-26 · measurement release: v1.7.8
Weights CC BY-NC-4.0 per the model card (non-commercial; listed for evaluation only); bases Apache-2.0; serving code Apache-2.0 (jeff 1.4.0).
intelligence
74.5
calibration
89.3
speed
74.0
cost
53.8
Composite (secondary)
70.5 · Composite rank #13
Cost and measurement conditions
$0.035 per 1,000 decisions (estimated)
ESTIMATE: explicit USD 0.05/0.15 per million input/output tokens, settled Gemma-4-12B estimate via google/gemma-3-12b-it at the full catalog pair. google/gemma-4-12B-it is unlisted in the frozen pricing module; no base_model key or exact-base floor applies. Native usage priced as reported; uncounted second-prefill estimate of 427,166 input tokens is disclosed and not priced. Common 1,485-item basis USD 0.03469 per 1,000 decisions = 0.537 times cap; all-1,500 basis USD 0.03758 = 0.582 times cap. Estimated, not charged usage or a host quote. If the uncounted prefills were priced as input (403,115 tokens on the common 1,485-item basis), the common-basis cost would be USD 0.048268 per 1,000 decisions, 0.747 times the cap (all-1,500 USD 0.051819, 0.802 times the cap). The Cost axis would fall from 53.79 to 49.49, below the 50-point composite gate. This sensitivity estimate is not included in the published price or scores.
Measured p50 latency: 0.067 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: 0.28 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.067 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 = 0.28 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
Measured 26 Oct 2026 on one rented NVIDIA RTX PRO 6000 Blackwell Server Edition 96 GB in Finland over loopback. One measured attempt in a fresh process with pinned weights and serving code, read-only model and source, no credentials and networking disabled. GPU rental is separate from the estimated model tariff.
Measured latency over 1,500 decisions: p50 0.067 s · p95 6.957 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. Adjusted p95 14.06 s, 0 decisions over the limit.