Instinct Dual 4Bv1.7 · 10 s rule

Data: JevBench v1.8.11, published 2026-11-15 · method v1.7 (10 s rule), re-scored from the recorded per-decision times

Capability 57.2 · API offering, ranked on the API leaderboard · Capability rank #33 of 39 within the Jev-class caps

Measured 2026-10-06 on A4 ∪ P, 600 items, equated (+2.57 I / +3.35 C); round score-a4-2.

Not independently recorded

Base model: Qwen3.5-4Bsource · Operator-reported; weights not publicly verifiable.

intelligence

28.3

calibration

86.1

speed

88.6

cost

75.0

Composite (secondary)

18.0 · Composite rank #44 of 54 on the API leaderboard

Cost and measurement conditions

$0.0068 per 1,000 decisions (public tariff)

measured: token usage of the 6 Oct 2026 A4 u P run x tariff USD 0.01/M input, 0/M output (spec instinct-dual-4b.json), per 1,000 decisions on the v1.6.1 common cost basis (591 of 600 items; the 9 A4 u P items outside the Jev reference's accepted input range are excluded, cost-basis-a4.json)

p50 latency: 0.2 s

none (API)

Measured latency over 600 decisions: p50 0.232 s · p95 0.601 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.

JevBench v1.6.43 values (method v1.6, before the 10 s rule)

Capability 57.21 · Composite 17.97 · rank #142 on the all-systems board · #36 API

intelligence 28.34 · calibration 86.08 · speed 88.60 · cost 75.03

v1.6 speed basis (124 items): p50 0.231 s · p95 0.598 s

These figures are kept for comparison. The current figures above are re-scored under the 10 s rule from the recorded per-decision times and replace them.

Breakdowns (not part of the score)

Chance-corrected competence (0 = chance, 100 = perfect) on A4 ∪ P (600 items: 300 open + 300 sealed; fewer sealed items than the full set, so wider uncertainty). Raw values, not equated. Category cells use A4+P+L1+L2+L3; per-type/tier cells use the original basis above. Radars against Jev.

Request typeOpenSealed
choice51.749.0
noul0.00.0
score27.126.7

Capability by subject topic

Rules, policy & law4.5n=1133
Coding & software31.6n=202
Math & numbers0.0n=269
Finance & commerce5.9n=178
Support & operations27.4n=109
Everyday language52.2n=130
Safety & security27.4n=84

Use cases (TypeSafe categories)

Model routing43.9n=267
Legal & compliance17.3n=269
Customer support12.3n=156
E-commerce4.0n=88
Insurance claims9.9n=92
Risk assessment7.2n=88
Financial crime0.7n=76
Feature extraction0.0n=90
Lead generation14.9n=86
Recruiting0.0n=88
Knowledge graphs0.0n=82
LLM guardrails6.7n=82
Moderation5.6n=69
Code linting0.0n=81
Search & retrieval35.4n=84
Science0.0n=82
Advertising0.0n=85
Gaming0.0n=79
Demand forecasting0.0n=72

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

Previous measurement (v1.5.x, not comparable)

Capability 57.2 · composite 18.0 · cost $0.0068 per 1,000 decisions (public list price)

Last measured 2026-10-06, measurement release v1.6.1. Older method and scale; the current figures above replace it.

Published source