Jev vs Winnow-12B Q8: published benchmark comparison

Data: JevBench v1.6.1, published 2026-10-06

Rank #13 on the open-weights board in v1.6.1.

Compare Jev-compatible models by separate measures and their measurement conditions. A historical measurement is not a current rank.

Jev and Winnow-12B Q8: published axes and category radars

Pick any two. Radars for the score axes, capability by subject topic and use cases, chance-corrected competence per request type on the open and sealed sets, and competence per tier on each set. Further out is better on every spoke; the link keeps the pair.

System A

System B

A: Winnow-12B Q8 — Open weights · LLM decoder · Score 68.9 (#13)
B: Jev 1.13.0REFERENCE · TypeSafe — Jev — reference (TypeSafe, closed) · Score 71.5 (ranked, not ranked)

The four score axes

0–100, the values in the table. An axis a system has no published value for is left as a gap (it counts as 0 in the composite).

Capability by subject topic

Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.

What the decision is about — math and dates, coding, rules, money, support work, everyday messages, safety. Chance-corrected competence per category (0 = at or below chance; negative averages are reported as 0, 100 = perfect), Winnow-12B Q8: S+P+L1+L2+L3; Jev 1.13.0: S+P+L1+L2+L3 items pooled; only categories with at least 30 items are spokes, smaller ones are listed below. Hover a category for its definition and item count.

What each category means · items per category

Rules, policy & law — applying written rules: company policies, contracts, regulations, eligibility. 1616 items (151 open / 1465 sealed)
Coding & software — code, SQL, repositories, developer tools and IT systems. 382 items (62 open / 320 sealed)
Math & numbers — a calculation decides the answer: arithmetic, word problems, probability, dates, units. 370 items (37 open / 333 sealed)
Finance & commerce — money: payments, refunds, invoices, orders, expenses, insurance payouts. 223 items (12 open / 211 sealed)
Support & operations — support tickets, incidents, logistics, scheduling desks and routing work to a team. 160 items (21 open / 139 sealed)
Everyday language — short everyday messages: intents, assistant requests, reading a detail out of a text. 145 items (11 open / 134 sealed)
Safety & security — untrusted or injected instructions, fraud, moderation, access and security triage. 109 items (6 open / 103 sealed)

Use cases (TypeSafe categories)

Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.

The real-world application area of each decision, after the TypeSafe use-case map; an item can count in two. Chance-corrected competence per category (0 = at or below chance; negative averages are reported as 0, 100 = perfect), Winnow-12B Q8: S+P+L1+L2+L3; Jev 1.13.0: S+P+L1+L2+L3 items pooled; only categories with at least 30 items are spokes, smaller ones are listed below. Hover a category for its definition and item count.A point on the bold 0 ring printed 0.0 is a measured value, not a gap: these published cells clip below-chance results to 0, so 0 means at or below chance. Their score markers cannot fall inside that ring.

What each category means · items per category

Model routing — deciding which LLM or model tier should handle a prompt; classifying a request's intent, domain, difficulty or risk to pick a model. 524 items (86 open / 438 sealed)
Legal and compliance — contracts, policies, regulations, eligibility rules, checking documents against written requirements. 515 items (74 open / 441 sealed)
Customer support — classifying tickets and customer intents, urgency, refunds, routing cases to a team, checking support replies. 240 items (37 open / 203 sealed)
Other — none of the above: general reasoning, arithmetic or date puzzles, everyday assistant requests without a business workflow. 174 items (24 open / 150 sealed)
E-commerce marketplaces — product listings, product attributes, orders and deliveries, seller catalog policy. 119 items (9 open / 110 sealed)
Insurance claims — first-notice-of-loss reports, claim complexity, missing information, claim triage. 119 items (9 open / 110 sealed)
Risk assessment — estimating probabilities and severity of risks, incidents, vendor or underwriting risk. 118 items (8 open / 110 sealed)
Financial crime — suspicious transactions, KYC, fraud alerts, entity matching for investigations. 101 items (9 open / 92 sealed)
Feature extraction for predictive modeling — turning natural-language data into probabilistic features or estimates for a downstream prediction. 99 items (5 open / 94 sealed)
Lead generation — matching companies or inbound messages to an ideal customer profile, buyer intent, prioritising and routing sales leads. 95 items (4 open / 91 sealed)
Recruiting — resumes, applications, interview feedback, matching candidates to roles. 94 items (1 open / 93 sealed)
Graphs and knowledge graphs — entity types, relationships between records, contradictions between facts, multi-step lookups across linked facts. 94 items (3 open / 91 sealed)
LLM guardrails — checking LLM inputs, outputs and tool calls: prompt injection, jailbreaks, policy violations, tool-call errors, answer quality. 93 items (5 open / 88 sealed)
Moderation and trust and safety — user content moderation: toxicity, harassment, spam, unsafe advice, personal data exposure. 93 items (7 open / 86 sealed)
Semantic code linting — checking code or writing against conventions and guidelines, as in CI review of code. 90 items (3 open / 87 sealed)
Search and retrieval — scoring query-to-candidate relevance, reranking results, selecting useful context for RAG. 90 items (5 open / 85 sealed)
Scientific discovery — screening papers, labelling research passages or survey answers, checking that citations support claims, methodology checks. 90 items (2 open / 88 sealed)
Advertising — ad creatives, campaign copy, brand safety, prohibited claims in ads. 88 items (3 open / 85 sealed)
Gaming — player reports, in-game chat, game support. 86 items (3 open / 83 sealed)
Demand forecasting — purchase intent, product interest, demand and supply signals for forecasting. 83 items (3 open / 80 sealed)

Competence per request type, open / sealed

Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.

Chance-corrected competence for Choice, Noul and Score on the 300 open and 1200 sealed decisions.

Competence per tier — open set

Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.

Per-tier competence, the three request types pooled by their published decision counts.

Competence per tier — sealed set

Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.

Per-tier competence on the sealed decisions, types pooled the same way; item text stays private.

How the categories were made

O1S chance-corrected competence: equal mean across the request types present, clipped to 0–100. Zero means the equal-weight mean across present request types was at or below baseline; negative means are clipped. It does not mean every type scored zero. Choice uses a random-option baseline; Noul uses a fixed 50% accuracy baseline (abstentions count as wrong); Score uses the midpoint-guess error (random-guess fallback when every gold is at the midpoint). Compare models within a category; row pools and supported types can differ.

Categories use ruled subject-topic labels and authoring use-case precedence across the answered pool union, with stable item identities counted once.

Official O1S; recorded failures count where supported; no missing response is imputed; stable identity dedupe parent first.
S-based category cells pool recorded supported responses from S+P and the answered L1/L2/L3 supplements. API overlay category cells keep their separately recorded A4/A5+P union. Headline and per-type/tier scores are unchanged.

All values as a table

Spoke	A: Winnow-12B Q8	B: Jev 1.13.0
The four score axes
Intelligence	59.5	63.6
Calibration	83.0	90.6
Speed	86.7	91.5
Cost	56.6	54.7
Capability by subject topic
Rules, policy & law	53.8	49.4
Coding & software	70.8	64.3
Math & numbers	14.8	10.4
Finance & commerce	51.7	40.9
Support & operations	56.4	48.8
Everyday language	74.2	77.5
Safety & security	72.6	52.2
Use cases (TypeSafe categories)
Model routing	74.3	73.0
Legal & compliance	61.2	68.6
Customer support	55.1	49.6
Other	25.6	40.1
E-commerce	45.4	40.5
Insurance claims	58.2	46.1
Risk assessment	38.3	32.8
Financial crime	36.9	35.8
Feature extraction	27.9	28.3
Lead generation	51.5	49.2
Recruiting	44.7	0.0
Knowledge graphs	51.8	28.8
LLM guardrails	82.7	68.0
Moderation	53.1	51.6
Code linting	62.3	29.1
Search & retrieval	84.1	80.9
Science	23.4	21.2
Advertising	33.7	16.4
Gaming	21.8	5.1
Demand forecasting	0.0	0.0
Competence per request type, open / sealed
Choice · open	70.5	79.0
Choice · sealed	69.7	75.8
Noul · open	50.5	54.5
Noul · sealed	52.5	46.1
Score · open	58.5	63.3
Score · sealed	55.5	63.0
Competence per tier — open set
Easy	86.9	93.9
Standard	57.3	63.2
Judge	52.1	67.7
Hard	65.1	65.6
Competence per tier — sealed set
Easy	77.2	78.4
Standard	65.7	68.7
Judge	68.3	66.9
Hard	52.3	58.9

Published values

Measure	Jev 1.13.0 (TypeSafe AI)	Winnow-12B Q8
Capability Score	77.1	71.2
Rank (open-weights board)	Reference, not ranked	#13
Composite (secondary)	71.5	68.9
intelligence	63.6	59.5
calibration	90.6	83.0
speed	91.5	86.7
cost	54.7	56.6
USD per 1,000 decisions	$0.032 per 1,000 decisions	$0.028 per 1,000 decisions
p50 latency (seconds)	0.2	0.117 measured on-GPU, no network · 0.38 est. incl. network (2x + 150 ms)
Last measured	2026-10-05	2026-10-02

* Self-hosted rows: the first value is the p50 we measured on our own hardware over loopback (on-GPU, or on-CPU where marked, no network). It is not directly comparable with API latency, which we time end to end from our measuring host, network included. “est.” is the est. incl. network (2x + 150 ms) that the ranking and the latency cap use for every self-hosted row. Model authors can ask us to time a hosted endpoint instead (on request).

Jev 1.13.0 (TypeSafe AI)

proprietary API

Published source

Measured on: the operator's hosted API
Cost basis (not published): TypeSafe list tariff USD 0.042 per 1M input tokens, output free; measured on the v1.6.1 common cost basis (all items except those outside the Jev reference input range) from this run's token usage
Latency adjustment: none (API)

Winnow-12B Q8

Apache-2.0, including the applicable Gemma 4 base/derivative licence terms

Published source

Measured on: evaluator-owned Lium GPU pod (RTX6000), offline read-only container
Cost basis (estimated): v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; no exact base-model floor applies
Latency adjustment: x2 + 0.15 s (assumption, not measured)

Frequently asked questions

What does JevBench show for Jev and Winnow-12B Q8?: Jev 1.13.0 has Capability 77.1 and is the unranked reference (the open-weights board ranks 92 open-weights systems; hosted API offerings are ranked on the API leaderboard). Winnow-12B Q8 has Capability 71.2. Rank #13 on the open-weights board in v1.6.1. Historical scores use their original method and are not ranked against the current release.
How does Capability differ from Composite?: Capability is the mean of Intelligence and Calibration; headline ranks require both official cost and median-latency caps. Composite additionally scores Speed and Cost and is secondary.
Can I compare the prices as actual bills?: Jev 1.13.0 (TypeSafe AI): not published, $0.032 per 1,000 decisions. Winnow-12B Q8: estimated, $0.028 per 1,000 decisions. Estimates and self-reported vendor prices are not measured charges.
Is Winnow-12B Q8 open source?: The published licence note for Winnow-12B Q8 is Apache-2.0, including the applicable Gemma 4 base/derivative licence terms. Jev 1.13.0 is listed as proprietary API. Check the linked sources for terms.