KnowLine 26B-A4B-Gen2.5v1.7 · 10 s rule
Data: JevBench v1.7.23, published 2026-10-29 · method v1.7 (10 s rule), measured under the rule
Capability 77.3 · Ranked outside the official caps (no Capability rank)
cost 2.9× Jev (v1.5 reference)
Last measured: 2026-10-23 · measurement release: v1.7.18
Apache-2.0 weights (model card and LICENSE); MIT author serving code (file headers); public, not gated
intelligence
65.2
calibration
89.4
speed
89.5
cost
40.9
Composite (secondary)
43.1 · Composite rank #67
Cost and measurement conditions
$0.093 per 1,000 decisions (estimated)
ESTIMATE (*est.), base-model market reference google/gemma-4-26b-a4b-it: USD 0.09 per million input tokens and USD 0.30 per million output tokens. Native server usage includes the chat template and read-twice layout; output is one label token per request. 1,531,517 input and 1,485 output tokens on the 1,485 common items. Cost on the 1,485 items every row on this draw answered (same-items rule); all-items cost USD 0.163650 per 1,000 decisions. Common-items cost USD 0.093119 per 1,000 decisions, 1.44 times the official cost cap. Estimated model tariff, not charged usage or a host quote; GPU rental is separate.
Measured p50 latency: 0.079 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: 0.31 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.079 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 = 0.31 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
Self-hosted on one evaluator-rented NVIDIA RTX PRO 6000 Blackwell Server Edition 96 GB GPU in Finland (EU), using the author's own server over SGLang 0.5.21. Contributor code ran in an offline container on a disposable rented GPU VM. Measured 23 Oct 2026. GPU rental is separate from the estimated model tariff.
Measured latency over 1,500 decisions: p50 0.079 s · p95 0.109 s
15 decisions over 10 s (15 of them answered without error, counted as wrong).
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.