whatmodel.sh / decision / Winnow-12B (DecisionBench b6ac22b0)

Winnow-12B (DecisionBench b6ac22b0)

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Winnow-12B (DecisionBench b6ac22b0) is an open-weights decision model from EldanRing. Licence: apache-2.0.

Catalogue identity notes

DecisionBench v1 comparison: EldanRing/Winnow-12B, revision b6ac22b0d51b69b18200acacb3fbdd98073fffe8; decision-model; adapter jev-compatible-server-native-systemone-grouped-v1; probability source native-with-anchored-wide-choice-and-noul-adaptation. Kept separate from newer or unpinned Decision Index recipes; standard prompt, not compact. No unverified OpenRouter price imported. Current repository metadata checked separately; an unpinned alias is not evidence that today’s weights reproduce this historical run. Reviewed Ollaya local route added 2026-10-31: winnow:12b, manifest 8d72fee0e5f92a625cdb85ae4f6abdc024c9e6f124a0ca67e64522c3f24c47ce, matching the independent accuracy-run manifest. Native family implementation; published runtime precision label Q8_0. Memory is measured peak GPU increase (15872 MiB, rounded upward to 15.5 GiB) on RTX 5090 during the short five-question triage workload, not download size or a maximum-context guarantee. Runtime installation is required; no fresh execution performed here. Historical benchmark adapter and input protocol still vary by result.

Model information sources

· Weights licence ↗

58.0±1.5#129 of 355 decision modelsfrom 41 benchmark results across 2 sources

$ how it compares

The decision models ranked either side of Winnow-12B (DecisionBench b6ac22b0), out of 357 with a score, then more from EldanRing.

#modelscorewhereRAMreleased
129Mica v0.5 4B akivet58±1model files2026-10
130AJev lora7 (Gemma 4 12B) andyzhang23258±1model files
131Winnow-12B (DecisionBench b6ac22b0) EldanRing58±2local~16GB
132deepseek-v4.1-flash (DecisionBench openrout) DeepSeek58±2model files
133Nace.AI Drex v1.5 Nace.AI58±1model files
more from EldanRing
163Winnow-12B EldanRing55±1model files
236Winnow-E4B EldanRing46±1model files

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

$ what it's good at

knowledge & reasoning▁▂▂▂▂▃▃▅▅▆█▂ #152 of 263 · 8.0% of its score · 5 tests

Trails Ryotide Qwen v0.3.1 (decision readout), Interfaze lev Qwen3.5-4B LoRA, razorback16 diffgemma (decision readout) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Ollaya · PubMedQA71.2% ✎▃▂▃▃·▂▃▃▆█▆▂ beats 13/25easy · avg 69%▇ 2.9%typical
Ollaya · BoolQ87.3% ✎▁··▃▃▁▄▁▃█▇▄ beats 20/25easy · avg 84%▇ 2.1%typical
Ollaya · SQuAD 2 answerability87.6% ✎▁▁▃▁▇█▃▃▄▃·▁ beats 24/25easy · avg 80%▇ 1.1%blunt
Ollaya · VitaminC verification75.5% ✎▁···▃·▄█▃··▁ beats 8/25easy · avg 78%▇ 1.1%blunt
Ollaya · MultiNLI81.6% ✎▂··▄██▂▆▆█▆▂ beats 6/25very easy · avg 87%▇ 0.7%blunt
workflows▁▁▁▂▁▂▂▆█▄▁▁ #37 of 114 · 91% of its score · 36 tests

Beats vLLM-SR d3-flash (9B), Clef-omni, Jebadiah 27B v2.1 (Ollaya Q4_K_M) here, though they rank higher overall. Trails StartLux-Decision v1.1 2B, Solar Decide, deepseek-v4.1-flash (DecisionBench openrout) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
DecisionBench v1 · action selection (all-request accuracy) submitted57.7% ✎▂▁▂▃▄▃▁▁▂█▇▄ beats 38/56medium · avg 51%▇▇▇▇▇▇ 15%at its level
DecisionBench v1 · routing (all-request accuracy) submitted77.0% ✎▁▁▁▁▂▂▁▁▁▁▃█ beats 52/56medium · avg 64%▇▇▇▇ 10%sharp · at its level
DecisionBench v1 · navigation (all-request accuracy) submitted73.0% ✎▁▁·▃▁▁▁▁▁▁▃█ beats 50/56medium · avg 58%▇▇▇ 7.9%sharp · at its level
DecisionBench v1 · record classification (all-request accuracy) submitted46.0% ✎▃▂▁▂▂▂▄▁▃▆██ beats 43/56medium · avg 39%▇▇▇ 7.2%at its level
DecisionBench v1 · verification (all-request accuracy) submitted57.5% ✎▁▁▅▁▃▂▂▁▁▇▇█ beats 29/56medium · avg 56%▇▇ 4.8%blunt · at its level
Ollaya · Typed decisions test70.2% ✎▄·▂▆▄▄▂·▄█▆▆ beats 17/22medium · avg 64%▇▇ 4.7%at its level
DecisionBench v1 · document record classification (all-request accuracy) submitted46.0% ✎▂▂·▁▂▄▃▆▄█▁▁ beats 32/56medium · avg 46%▇▇ 4.0%at its level
DecisionBench v1 · rubric scoring prioritization (all-request accuracy) submitted56.1% ✎▁▁▅▅▃▃▅█▃▃·▁ beats 49/56medium · avg 45%▇ 3.4%at its level
28 more tests
testresultvs the others in its rundifficultycountswhy
DecisionBench v1 · triage (all-request accuracy) submitted84.0% ✎▁·▁·▂▃▄▃▂█▄▂ beats 35/56easy · avg 78%▇ 3.0%sharp
Ollaya · SummEval relevance49.6% ✎▁··▁▂▁▅··▃██ beats 20/25medium · avg 41%▇ 2.9%at its level
DecisionBench v1 · completion assessment (all-request accuracy) submitted93.0% ✎▁▁▂▃▄·▂▁▁▆█▂ beats 47/56very easy · avg 86%▇ 2.7%sharp · maxed out
Ollaya · PAWS paraphrase83.2% ✎▂▂▂▂·▃█▆·▂·▁ beats 19/25easy · avg 81%▇ 2.6%typical
DecisionBench v1 · content classification (all-request accuracy) submitted79.0% ✎▁····▁▁▁▁▁▁█ beats 40/56easy · avg 73%▇ 2.3%typical
Ollaya · Aegis 2 safety81.6% ✎▂·▁▁▂·▃▃▂█▃▁ beats 21/25easy · avg 77%▇ 2.2%typical
DecisionBench v1 · review scoring (all-request accuracy) submitted62.0% ✎▁······▂▃▅▇█ beats 52/56medium · avg 52%▇ 2.2%at its level
DecisionBench v1 · bounded extraction (all-request accuracy) submitted94.5% ✎▃▃▂·▁▁▁▁▁▁█▅ beats 44/56very easy · avg 86%▇ 2.2%sharp · maxed out · noisy
Ollaya · MASSIVE German intent85.1% ✎▁····▁·▂▃▆█▃ beats 16/23easy · avg 84%▇ 1.9%sharp
DecisionBench v1 · retrieval ranking (all-request accuracy) submitted57.0% ✎▁▁·▁▁▂▃▂▃▁▁█ beats 43/56medium · avg 35%▇ 1.8%at its level · noisy
DecisionBench v1 · target selection (all-request accuracy) submitted50.0% ✎▁▁·▁▁▁▃▁▂▁▁█ beats 56/56hard · avg 29%▇ 1.7%at its level · noisy
DecisionBench v1 · risk scoring (all-request accuracy) submitted64.0% ✎▁▁·▁▂▂▁▃▃█▇▃ beats 34/56medium · avg 62%▇ 1.5%blunt · at its level
Ollaya · SummEval consistency87.5% ✎▁·▁▁··▁▁▁▁▄█ beats 25/25easy · avg 75%▇ 1.1%sharp · noisy
Ollaya · HelpSteer 2 scoring40.6% ✎▁▁··▁·▁▃▇█▅▄ beats 17/23medium · avg 38%▇ 1.1%blunt · at its level
Ollaya · MASSIVE English intent86.6% ✎▁·······▂▃█▃ beats 19/23very easy · avg 86%▇ 1.0%noisy
DecisionBench v1 · reasoning (all-request accuracy) submitted70.3% ✎▂·▁▁▆█▅▃▃▂▁▁ beats 49/56medium · avg 52%▇ 0.9%at its level
DecisionBench v1 · document workflows (all-request accuracy) submitted84.5% ✎█▄▃▁▁▃▂▃▁·▁▁ beats 52/56easy · avg 66%▇ 0.6%blunt · at its level
DecisionBench v1 · retrieval verification (all-request accuracy) submitted87.5% ✎█▁▂▁▁▁▁▁▂▅▁▁ beats 44/56easy · avg 74%▇ 0.6%noisy
DecisionBench v1 · guardrails moderation (all-request accuracy) submitted63.6% ✎▁▁▁▂▆▃▃█···▁ beats 53/56medium · avg 52%▇ 0.4%blunt · at its level · noisy
DecisionBench v1 · safety gating (all-request accuracy) submitted96.0% ✎▂·····▆█▂▂▄▇ beats 47/56easy · avg 77%▇ 0.3%noisy
DecisionBench v1 · document classification (all-request accuracy) submitted38.0% ✎▂▁·▂▁▁▁▄▁▂▂█ beats 54/56hard · avg 21%▇ 0.3%blunt · noisy
DecisionBench v1 · semantic filtering (all-request accuracy) submitted86.0% ✎▁·▁▁·▁▁█▂▃█▂ beats 51/56easy · avg 71%▇ 0.2%blunt · noisy
Ollaya · Civil Comments87.0% ✎▂▂▂·▂▅▅▆▃▅·█ beats 13/23very easy · avg 86%▇ 0.1%blunt · noisy
DecisionBench v1 · content moderation (all-request accuracy) submitted95.0% ✎▁····▃▅▃▂▃·█ beats 36/56easy · avg 82%▇ 0.1%noisy
DecisionBench v1 · context management (all-request accuracy) submitted90.5% ✎▁·····▁▇▆▁█▃ beats 46/56easy · avg 73%▇ 0.1%blunt · noisy
DecisionBench v1 · entity alignment (all-request accuracy) submitted99.5% ✎▁·▁▄▇▄▅▄▂▁▄█ beats 56/56easy · avg 66%▇ 0.1%at its level · noisy
DecisionBench v1 · function agent skill routing (all-request accuracy) submitted95.9% ✎▃▃▄▅▆▂▅▄▃▄█▃ beats 43/56easy · avg 68%▇ 0.1%blunt · at its level · noisy
DecisionBench v1 · routing triage (all-request accuracy) submitted75.5% ✎█▁···▁·▂▆▅▂▂ beats 48/56medium · avg 42%▇ 0.0%blunt · at its level · noisy

$ where to run it

Other ways to run it

  1. Run it as published · needs 15.5 GB, NVIDIA only · from github.com/ollaya-dev
    ollaya run winnow:12b --preset triage "The customer requests a refund."

Download model files ↗
Published model files verified. This is the revision recorded by the benchmark. A matching Ollaya runtime manifest and published short-workload GPU memory measurement were reviewed separately; see local variants and identity_note. The command was not executed in this review. Source · checked 2026-10-31.

Model repository and documentation: huggingface.co/EldanRing ↗

$ how the score was worked out · Hanno Labs DecisionBench v1 r… 73%, Ollaya · original runs 26%

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

sourceresultsshare of the scorelicence
Hanno Labs DecisionBench v1 reviewed results · submitted; independent execution unverified27CC0 1.0
Ollaya · original runs14Apache-2.0
prior · pull toward the middle
weighted average= 58.0

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit 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 · sources and licences · json