Winnow-12B

+ contribute

Winnow-12B is an open-weights decision model from EldanRing. Its published size is 11.96B parameters. Licence: apache-2.0.

Catalogue identity notes

Exact Decision Index tested engine winnow-12b-q8; reported model Winnow-12B. Type: LoRA; technique: autoregressive. Published checkpoint/base reference: EldanRing/Winnow-12B. Implementation: https://github.com/EldanRing/winnow-inference. Comparison entry; no current run command, API price or speed inferred from the benchmark. Current repository metadata checked separately; an unpinned alias is not evidence that today’s weights reproduce this historical run.

Model information sources

· Weights licence ↗ · Parameter count ↗

Publisher parameter count; includes the served backbone, not just adapter size.

54.7±1.2#161 of 355 decision modelsfrom 58 benchmark results across 3 sources

$ how it compares

The decision models ranked either side of Winnow-12B, out of 357 with a score, then more from EldanRing.

#	model	score	where	RAM
161	Standard One 8B SH (v4) Standard Thinking	55±1	model files
162	kev-8b (DecisionBench c80773da) Jared Palmer	55±2	model files
163	Winnow-12B EldanRing	55±1	model files
164	ezjev 4B s2 everettjf	55±1	model files
165	JoshuaSP diffgemma (decision readout) JoshuaSP	55±1	source code
	more from EldanRing
131	Winnow-12B (DecisionBench b6ac22b0) EldanRing	58±2	local	~16GB
236	Winnow-E4B EldanRing	46±1	model files

see all decision models · find the ones that fit your machine

$ what it's good at

tools & automation▁▃▁▁▁▂▃▄▆██▆ #125 of 250 · 6.4% of its score · 6 tests

Beats Tern-35B-A3B, GEV-26B-Decide, Cerno 4B v1.1 here, though they rank higher overall. Trails OneJev1.6-0.8B, TokenRhythm NeoHorse-Jev-4B, razorback16 diffgemma (decision readout) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Decision Index 0.4.1: API-Bank (accuracy) submitted	67.3% ✎	▁▁▁▁▁▁▁▂▃▃█▂ beats 85/236	easy · avg 70%	▇ 2.0%	sharp · noisy
Decision Index 0.4.1: BFCL (case exact accuracy) submitted	94.8% ✎	▁▂▁▁▁▂▂▄▅▆█▁ beats 125/246	very easy · avg 92%	▇ 1.7%	sharp · maxed out · noisy
Decision Index 0.4.1: Home appliance simulator (case exact accuracy) submitted	54.5% ✎	█▁▁▃▄▄▆▄▁▂▁▃ beats 140/243	medium · avg 37%	▇ 1.0%	sharp · at its level · noisy
Decision Index 0.4.1: When2Call (accuracy) submitted	83.2% ✎	▁▁▁▂▃▅▅▆█▇█▃ beats 164/246	easy · avg 72%	▇ 0.9%	noisy
Decision Index 0.4.1: ToolRet (nDCG@10) submitted	65.2% ✎	▁▁▁▁▁·▁▁▁▄█▁ beats 140/246	medium · avg 60%	▇ 0.8%	blunt · at its level
Decision Index 0.4.1: RouterBench (selected quality (quality objective)) submitted	79.8% ✎	no other model yet		—	failed run: Publisher excludes this protocol because its prompt reveals GPT-4 as the best route. Evidence: https://huggingface.co/spaces/multimodalart/jev-decision-index/blob/c130faf65f82e81d3bdefb01e758c712dc5c5199/data/index.json

retrieval & classification▁▁▁·▁▁▁▁▃▃█▅ #125 of 258 · 5.6% of its score · 7 tests

Beats AutoTrust-JEV-27B v1.1, Quyet-1.0-Large, Tern-35B-A3B here, though they rank higher overall. Trails vLLM-SR vela3-fast (Vela 3.0, 0.4B), crystal-0.8b, Kas 4B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Decision Index 0.4.1: BRIGHT (nDCG@10) submitted	45.3% ✎	▁▁▁▁▁▁▁▂▂▄▇█ beats 158/246	medium · avg 38%	▇ 2.0%	blunt · at its level
Decision Index 0.4.1: Amazon ESCI (macro-F1) submitted	53.0% ✎	▁▁▁▁▁▁▂▂▃▃█▅ beats 129/246	medium · avg 47%	▇ 1.9%	at its level
Decision Index 0.4.1: HoVer (accuracy) submitted	72.0% ✎	▆▆▇▅▄▄▃▇▄█▆▄ beats 109/246	easy · avg 74%	▇ 0.5%	blunt
Decision Index 0.4.1: PhishNChips phishing decisions (accuracy) submitted	65.9% ✎	█▅▅▄▃▂▁▁▁▁·▁ beats 110/245	easy · avg 70%	▇ 0.5%	at its level · noisy
Decision Index 0.4.1: CLINC150+OOS (macro-F1) submitted	87.0% ✎	▁▁▁▁▁▁▁▂▃▆█▄ beats 92/238	easy · avg 83%	▇ 0.4%	sharp · noisy
Decision Index 0.4.1: BANKING77 (macro-F1) submitted	75.1% ✎	▁·▁▁▁▁▁▁▂█▇█ beats 64/238	easy · avg 79%	▇ 0.3%	noisy
Decision Index 0.4.1: SGD/SGD-X (macro-F1) submitted	11.9% ✎	no other model yet		—	failed run: Publisher reports a builder bug exposing dialogue turns after the classified turn; awaiting corrected rerun. Evidence: https://huggingface.co/spaces/multimodalart/jev-decision-index/blob/c130faf65f82e81d3bdefb01e758c712dc5c5199/data/index.json

language understanding▂▂▂▃▂▅▆▃█▆▄▃ #139 of 256 · 20% of its score · 9 tests

Beats Solar Decide here, though they rank higher overall. Trails Ryotide Qwen v0.3.1 (decision readout), Sieve 2B v3, RSI-Jev v5.0 VL 3B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Decision Index 0.4.1: NLI4CT (macro-F1) submitted	82.9% ✎	▁▁▁▂▂▃▃▄▇▇█▄ beats 186/246	easy · avg 74%	▇▇ 4.6%	typical
Decision Index 0.4.1: HellaSwag (accuracy) submitted	85.1% ✎	▂▄▂▂▄▇▃▇█▅▃▁ beats 114/246	very easy · avg 85%	▇▇ 4.6%	sharp
Decision Index 0.4.1: ANLI (macro-F1) submitted	66.7% ✎	▁▂▃▅▅█▂····▁ beats 137/246	medium · avg 60%	▇▇ 4.5%	at its level
Decision Index 0.4.1: iSarcasmEval (Sarcasm F1 · track A, English) submitted	57.7% ✎	▁·▁·▁▁▂▄▅█▇▄ beats 173/246	medium · avg 44%	▇ 2.0%	at its level
Decision Index 0.4.1: ACOS (per-review F1) submitted	15.5% ✎	▁··▂▃▃▅▇█▇▃▃ beats 81/246	hard · avg 22%	▇ 1.0%	at its level · noisy
Decision Index 0.4.1: ContractNLI (macro-F1) submitted	72.9% ✎	▁▁▁▁▁▁▁▂▂▇▇█ beats 94/246	easy · avg 71%	▇ 0.9%	noisy
Decision Index 0.4.1: VAST (macro-F1) submitted	68.6% ✎	▁▁▁▂▅▄▄▂▄▆█▄ beats 129/246	medium · avg 62%	▇ 0.9%	at its level
Decision Index 0.4.1: RAGTruth (F1 on hallucinated class) submitted	77.1% ✎	▁▁▁▁▁▁▁▁▁▃▆█ beats 145/246	easy · avg 66%	▇ 0.7%	at its level · noisy

1 more test

test	result	vs the others in its run	difficulty	counts	why
Decision Index 0.4.1: FinEntity (macro-F1) submitted	91.7% ✎	▁···▁▁▁▁▃▆█▃ beats 173/246	easy · avg 84%	▇ 0.6%	noisy

arts & human taste▁▇▄▃▅▅█▅▄▂▁▁ #146 of 247 · 8.0% of its score · 6 tests

Beats OneJev1.6-4B, Nace.AI Drex v1.5, Jegue 0.1 (9B) here, though they rank higher overall. Trails Crivo-32M v2, Crivo-68M, Crivo-150M v2 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Decision Index 0.4.1: New Yorker (accuracy) submitted	64.8% ✎	▁▁▂▃▃▃▇▅█▅▃▁ beats 127/246	medium · avg 61%	▇ 3.4%	at its level
Decision Index 0.4.1: Humicroedit (accuracy) submitted	61.3% ✎	▁·▁▃▂▂▅█▄▂▁▁ beats 127/246	medium · avg 60%	▇ 1.6%	blunt · at its level
Decision Index 0.4.1: BPoMP (accuracy) submitted	83.1% ✎	▁█▄▄▅█▆▅▄▁▁▁ beats 100/246	easy · avg 83%	▇ 1.0%	sharp · noisy
Decision Index 0.4.1: cfcolor (accuracy) submitted	60.2% ✎	▁········▁▇█ beats 103/246	medium · avg 60%	▇ 0.9%	blunt · at its level
Decision Index 0.4.1: POP909-CL (accuracy) submitted	8.1% ✎	▄▅▅▃▆▅▃▅▂▅▃█ beats 93/237	hard · avg 21%	▇ 0.7%	sharp · at its level · noisy
Decision Index 0.4.1: Habermas Machine (accuracy) submitted	41.8% ✎	▁▂▄█▇▅▁▁▄▄▁▁ beats 92/246	medium · avg 47%	▇ 0.3%	blunt · at its level

knowledge & reasoning▁▂▂▂▂▃▃▅▅▆█▂ #120 of 263 · 39% of its score · 14 tests

Beats OneJev1.6-4B, Winnow-12B (DecisionBench b6ac22b0), Cerno 4B v1.1 here, though they rank higher overall. Trails PurpleMIST-Flash-1.0, Jevfire Qwen3.8-27B (decision readout), JevK5 9B v0.3.3 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Decision Index 0.4.1: MMLU-Pro (accuracy) submitted	54.8% ✎	▄▄▃▃▂▄▅▇██▅▁ beats 134/246	medium · avg 46%	▇▇▇▇ 8.9%	sharp · at its level
Decision Index 0.4.1: MMLU (accuracy) submitted	77.3% ✎	▁▄▄▃▃▂▅▇▅█▇▁ beats 134/246	easy · avg 70%	▇▇ 6.1%	sharp
Decision Index 0.4.1: BBH (accuracy) submitted	71.3% ✎	▄▆▄▄█▇▇▄▃▁▁▁ beats 140/246	easy · avg 66%	▇▇ 5.5%	at its level
Decision Index 0.4.1: ARC-Challenge (accuracy) submitted	94.8% ✎	▂▂▂▂▂▂▁▃▄▆█▆ beats 139/246	very easy · avg 90%	▇ 3.5%	sharp · maxed out
Decision Index 0.4.1: GSM8K (accuracy (4- and 10-choice)) submitted	48.6% ✎	▁··▄█▆▆▆▅▅▃▁ beats 136/246	medium · avg 45%	▇ 3.0%	sharp · at its level
Decision Index 0.4.1: WinoGrande (accuracy) submitted	74.1% ✎	█▅█▅▄█▆▄▃▃▁▁ beats 103/246	easy · avg 78%	▇ 2.7%	typical
Decision Index 0.4.1: GPQA Diamond (accuracy) submitted	38.8% ✎	▂▆▅▅▆█▆▃▁··▁ beats 110/246	medium · avg 39%	▇ 2.7%	blunt · at its level
Decision Index 0.4.1: ARC-Easy (accuracy) submitted	98.0% ✎	▁▁▂▁▂▂▂▂▃▆█▅ beats 128/246	very easy · avg 96%	▇ 1.8%	sharp · maxed out · noisy

6 more tests

test	result	vs the others in its run	difficulty	counts	why
Decision Index 0.4.1: SATA-Bench (case exact accuracy) submitted	33.6% ✎	▁▁▁▁▁▁▁▂▅█▂▁ beats 189/246	hard · avg 22%	▇ 1.2%	at its level
Decision Index 0.4.1: MuSR (accuracy) submitted	61.0% ✎	▁▁▂█▅▄▁▁▁▁▁▁ beats 144/246	medium · avg 60%	▇ 1.0%	blunt · at its level
Decision Index 0.4.1: CRUXEval (accuracy) submitted	69.5% ✎	██▄▃▃▄▁▁···▁ beats 168/246	medium · avg 58%	▇ 0.9%	at its level · noisy
Decision Index 0.4.1: CLadder (accuracy) submitted	68.6% ✎	▁███▆▅▃▃▁▁▁▁ beats 121/246	easy · avg 71%	▇ 0.8%	at its level
Decision Index 0.4.1: ChessBench (accuracy) submitted	17.3% ✎	▄▇▄▄▄▆▄█▄▁▁▁ beats 138/246	hard · avg 16%	▇ 0.6%	blunt
Decision Index 0.4.1: HLE (accuracy) submitted	10.4% ✎	▁▁▂▃▆█▇▅▅▃▁▁ beats 46/246	very hard · avg 13%	▇ 0.0%	blunt

vision▁·▃▂▃▃▇▄▄█▄▃ #51 of 79 · 12% of its score · 14 tests

Trails JPT-4B Qwen3.5-4B LoRA, LiquidAI d1-3B, JPT-9B Qwen3.5-9B LoRA here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Decision Index Vision 0.4 · Mind2Web (chance-corrected points) submitted	76.1600 points ✎	▁·▁▁▃▁▄▄▃▅█▂ beats 32/61	avg 63.423 points	▇ 1.3%	blunt
Decision Index Vision 0.4 · MMMU-Pro vision (chance-corrected points) submitted	27.6700 points ✎	▁▁▅▁▄▁▄█▄▃▇▇ beats 22/61	avg 22.797 points	▇ 1.2%	blunt
Decision Index Vision 0.4 · Winoground (chance-corrected points) submitted	80.7500 points ✎	▁·▁▁▁▁·▂▁▃▅█ beats 35/61	avg 66.987 points	▇ 1.1%	blunt
Decision Index Vision 0.4: private same skills (points)	53.6400 points ✎	▁▁▃▃▁▂▄▆▃▃▅█ beats 29/61	avg 48.067 points	▇ 1.1%	blunt
Decision Index Vision 0.4 · Moderation (Hateful Memes) (chance-corrected points) submitted	31.5700 points ✎	▆▃▃▆▄▄▇██▁▆▆ beats 24/61	avg 28.443 points	▇ 1.0%	blunt
Decision Index Vision 0.4 · RealWorldQA (chance-corrected points) submitted	52.8300 points ✎	▁·▂▂▁▃▃▅▅▃▂█ beats 14/61	avg 53.651 points	▇ 0.9%	blunt
Decision Index Vision 0.4 · CharXiv (chance-corrected points) submitted	67.4200 points ✎	▂▁▁▁▁▃▁▅▅▆██ beats 23/61	avg 62.137 points	▇ 0.9%	blunt
Benchmark Heaven imagejevbench v0.3.0: intelligence (points)	38.6319 points ✎	▃·▃▃··▅█▅▅·▃ beats 7/12	avg 32.235 points	▇ 0.8%	blunt

6 more tests

test	result	vs the others in its run	difficulty	counts	why
Decision Index Vision 0.4 · InfographicVQA (chance-corrected points) submitted	85.2500 points ✎	▁·▁▂▁▁▃▃▂▂▅█ beats 21/61	avg 83.859 points	▇ 0.7%	blunt
Decision Index Vision 0.4 · R-Bench-M (chance-corrected points) submitted	24.9200 points ✎	▂▄▇▆█▄▄▇▂··▁ beats 35/61	avg 19.676 points	▇ 0.7%	blunt
Decision Index Vision 0.4 · BLINK (chance-corrected points) submitted	51.1900 points ✎	▁···▁▂▂▂█▄▂▃ beats 21/61	avg 44.994 points	▇ 0.6%	blunt
Decision Index Vision 0.4 · CV-Bench (chance-corrected points) submitted	67.5400 points ✎	▁▁▁▁▂▁▃▁▄█▃▃ beats 13/61	avg 67.662 points	▇ 0.5%	blunt
Decision Index Vision 0.4 · KIE (CORD+FUNSD) (chance-corrected points) submitted	90.5800 points ✎	▁··▁▁▁▁▁▂▄█▁ beats 26/61	avg 85.184 points	▇ 0.4%	blunt
Benchmark Heaven imagejevbench v0.3.0: calibration (points)	61.3546 points ✎	▃···▃·▅█▅▃▃▅ beats 1/12	avg 68.584 points	▇ 0.3%	blunt

other tests4 not in a task · 10% of its score

test	result	vs the others in its run	difficulty	counts	why
Decision Index 0.4.1: private same skills (equated points)	47.6400 points ✎	▂▂▂▃▃▃▄▅▅▇█▃ beats 132/247	avg 41.206 points	▇▇ 4.5%	blunt
Decision Index 0.4.1: private new domains (equated points)	44.9500 points ✎	▃█▃▃▂▆▃▃▄▄▄▄ beats 144/247	avg 38.861 points	▇ 2.5%	blunt
Benchmark Heaven jevbench v1.6.1: intelligence (points)	59.5337 points ✎	▃···▃·▃▃█▅▅█ beats 7/13	avg 32.166 points	▇ 2.0%	blunt
Benchmark Heaven jevbench v1.6.1: calibration (points)	82.9581 points ✎	▂······▂▆▃██ beats 5/13	avg 70.133 points	▇ 0.6%	blunt

$ where to run it

Download model files ↗
Published model files verified. This is a verified repository snapshot; the benchmark does not establish an identical weight revision. A local run command and hardware requirements have not been verified for this entry. Source · checked 2026-10-31.

Model repository and documentation: huggingface.co/EldanRing ↗

$ how the score was worked out · Decision Index 0.4.1 85%, Decision Index Vision 0.4 10%, Benchmark Heaven (own decisio… 4%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
Decision Index 0.4.1 · submitted; independent execution unverified	42	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 85%	CC BY 4.0
Decision Index Vision 0.4 · submitted; independent execution unverified	12	▇▇▇ 10%	CC BY 4.0
Benchmark Heaven (own decision evaluations)	4	▇ 4%	CC BY 4.0
prior · pull toward the middle		▇ 1%
weighted average		= 54.7

not counted (2)

Decision Index 0.4.1	Decision Index 0.4.1: RouterBench (selected quality (quality objective)) [correct]	79.8%	failed run: Publisher excludes this protocol because its prompt reveals GPT-4 as the best route. Evidence: https://huggingface.co/spaces/multimodalart/jev-decision-index/blob/c130faf65f82e81d3bdefb01e758c712dc5c5199/data/index.json
Decision Index 0.4.1	Decision Index 0.4.1: SGD/SGD-X (macro-F1) [correct]	11.9%	failed run: Publisher reports a builder bug exposing dialogue turns after the classified turn; awaiting corrected rerun. Evidence: https://huggingface.co/spaces/multimodalart/jev-decision-index/blob/c130faf65f82e81d3bdefb01e758c712dc5c5199/data/index.json

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-11-03 02:37 UTC · sources and licences · json