GLiClass v4 9B (Knowledgator)v1.7 · 10 s rule
Data: JevBench v1.7.9, published 2026-10-26 · method v1.7 (10 s rule), measured under the rule
Capability 63.2 · rank #59 of 168 on the open-weights board
Last measured: 2026-10-26 · measurement release: v1.7.9
Apache-2.0 per the original model card (weights); the Apex republish includes the canonical Apache-2.0 LICENSE text; the original pinned repository has no separate LICENSE file; gliclass wheel Apache-2.0; base Qwen/Qwen3.5-9B Apache-2.0
intelligence
39.6
calibration
86.8
speed
85.6
cost
51.2
Composite (secondary)
37.0 · Composite rank #77
Cost and measurement conditions
$0.042 per 1,000 decisions (estimated)
ESTIMATE from base Qwen/Qwen3.5-9B at USD 0.10/0.15 per million input/output tokens; output priced at zero because this classifier generates no text. Input is the joint sequence actually encoded, including instructions and option labels: 626,409 native input tokens across 1,485 answered items, independently verified with zero mismatches; zero output tokens. The 15 HTTP 422 length refusals carry zero usage and count as misses. Common 1,485-item basis USD 0.04218 per 1,000 decisions = 0.653 times cap; all-1,500 basis USD 0.04176 = 0.647 times cap. Estimated usage price, not charged usage or a host quote.
Measured p50 latency: 0.135 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: 0.42 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.135 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 = 0.42 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 A6000 48 GB in Finland, over loopback. One measured attempt with pinned weights and serving code in a fresh process, 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 0.135 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.