Turion Jev 31Bv1.7 · 10 s rule
Data: JevBench v1.7.1, published 2026-10-25 · method v1.7 (10 s rule), measured under the rule
Capability 80.9 · Not ranked in the current release
cost 3.0× Jev (v1.5 reference)
Last measured: 2026-10-24 · measurement release: v1.7.1
Apache-2.0 weights (model card, LICENSE and NOTICE; the LICENSE appendix placeholder is unfilled); author serving code under the same repository licence (Apache-2.0 per the card and the pyproject.toml licence field; the six author modules carry no per-file licence headers); public, not gated
intelligence
70.9
calibration
90.9
speed
85.9
cost
40.6
Composite (secondary)
42.9 · Composite rank #59
Cost and measurement conditions
$0.096 per 1,000 decisions (estimated)
Estimated base-model market reference, not charged. Frozen pricing module aa2afcd5 resolves the author's exact google/gemma-4-31B-it spelling to USD 0.09/M input and USD 0.34/M output. Usage is 1580092 input tokens including the author's read-twice layout and zero reported output tokens. Zero-output accounting follows the measured native usage convention. Common basis excludes the same 15 long items for all token-priced rows. Cost exceeds the official cap; included in uncapped views and excluded from official cost-capped views. No cap exception. Cost on the common basis: USD 0.095763151515 per 1000 decisions; official cap USD 0.06459465517241379.
Measured p50 latency: 0.124 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: 0.40 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.124 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 = 0.40 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
Author's turion_jev server on evaluator-rented Verda RTX PRO 6000 Blackwell 96 GB, Finland, EU, 24 October 2026 19:12-19:31 UTC. Pinned local weights and code, isolated container, serial native requests. Shred and exact-instance teardown receipts retained. Approximately USD 0.83 rental is separate from the estimated inference tariff.
Measured latency over 1,500 decisions: p50 0.124 s · p95 0.251 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.
Model and serving disclosure
First measurement of turion-ai/turion-jev-31b at 9edd5783, a Gemma-4-31B-it fine-tune. Author's turion_jev server, in-process vLLM 0.32.0 with the author's official online FP8 engine options; evaluator launcher and score re-key shim. FP8 versus BF16 agreement is unknown. Training data partly comes from a Jev 1.13-distilled stream; contamination was not checked. No LFS weight shards are shared with KnowLine-31B-Gen1 or the base. Apache-2.0 LICENSE appendix is unfilled. Exactly 1485 answers and 15 native HTTP 422 refusals under the author's 65536-token maximum; the refused items count wrong in the full 1500-item denominator. Native usage counts the author's read-twice layout and reports zero output tokens. First 50 responses repeated bit-identically. No evaluator tuning or second measurement was performed.