whatmodel.sh / decision / mmastrac diffgemma (decision readout)

mmastrac diffgemma (decision readout)

+ contribute

mmastrac diffgemma (decision readout) is a source-available decision implementation from vllm-project. Its published size is 25.82B parameters.

Catalogue identity notes

Exact Decision Index tested engine vllm-pr57250; reported model mmastrac diffgemma. Type: inference technique; technique: diffusion. Published checkpoint/base reference: google/diffusiongemma-26B-A4B-it. Implementation: https://github.com/vllm-project/vllm/pull/57250. Comparison entry; no current run command, API price or speed inferred from the benchmark.

Model information sources

· Parameter count ↗

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

48.4±1.1#219 of 355 decision modelsfrom 50 benchmark results across 2 sources

$ how it compares

The decision models ranked either side of mmastrac diffgemma (decision readout), out of 357 with a score, then more from vllm-project.

#	model	score	where	RAM	released
219	Kev 4B r10 Jared Palmer	49±1	local	~11GB	2026-09-24
220	StartLux-Decision v1.1 0.8B StartLux	49±1	model files
221	mmastrac diffgemma (decision readout) vllm-project	48±1	source code
222	InternLM Intern-Decision Qwen3.5-4B Shanghai AI Laboratory (InternLM)	48±1	model files
223	Solar Decide Flash Upstage	48±1	api price unknown
	more from vllm-project
224	razorback16 diffgemma (decision readout) vllm-project	48±1	source code

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

$ what it's good at

tools & automation▁▃▁▁▁▂▃▄▆██▆ #98 of 250 · 5.4% of its score · 5 tests

Beats Jev 1.13, Xor 26B-A4B v1.2, Senova-4B v2 here, though they rank higher overall. Trails OneJev1.6-0.8B, TokenRhythm NeoHorse-Jev-4B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Decision Index 0.4.1: BFCL (case exact accuracy) submitted	96.2% ✎	▁▂▁▁▁▂▂▄▅▆█▁ beats 157/246	very easy · avg 92%	▇ 2.6%	sharp · noisy
Decision Index 0.4.1: When2Call (accuracy) submitted	48.7% ✎	▁▁▁▂▃▅▅▆█▇█▃ beats 39/246	easy · avg 72%	▇ 1.1%	at its level · noisy
Decision Index 0.4.1: Home appliance simulator (case exact accuracy) submitted	0.0% ✎	█▁▁▃▄▄▆▄▁▂▁▃ beats 46/243	medium · avg 37%	▇ 0.9%	sharp · at its level · noisy
Decision Index 0.4.1: ToolRet (nDCG@10) submitted	46.3% ✎	▁▁▁▁▁·▁▃▁▄█▁ beats 23/246	medium · avg 60%	▇ 0.8%	blunt · at its level
Decision Index 0.4.1: RouterBench (selected quality (quality objective)) submitted	79.9% ✎	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▁▁▁·▁▁▁▁▃▃█▅ #128 of 258 · 4.7% of its score · 5 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	22.6% ✎	▁▁▁▁▁▃▁▂▂▄▇█ beats 25/246	medium · avg 38%	▇ 1.9%	blunt · at its level
Decision Index 0.4.1: Amazon ESCI (macro-F1) submitted	53.4% ✎	▁▁▁▁▁▁▂▂▃▃█▅ beats 136/246	medium · avg 47%	▇ 1.8%	at its level
Decision Index 0.4.1: HoVer (accuracy) submitted	69.1% ✎	▆▆▇▅▄▄▃▇▄█▆▄ beats 95/246	easy · avg 74%	▇ 0.5%	blunt · at its level
Decision Index 0.4.1: PhishNChips phishing decisions (accuracy) submitted	82.4% ✎	█▅▅▄▃▂▁▁▁▁·▁ beats 183/245	easy · avg 70%	▇ 0.5%	at its level · noisy
Decision Index 0.4.1: SGD/SGD-X (macro-F1) submitted	43.0% ✎	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▂▂▂▃▂▅▆▃█▆▄▃ #163 of 256 · 22% of its score · 9 tests

Beats PurpleMIST-Flash-1.0, Hopper 12B (trained), Jev-Omni here, though they rank higher overall. Trails DecisionTune 1.0, Decider 2B, vLLM-SR d3-lite (0.8B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Decision Index 0.4.1: HellaSwag (accuracy) submitted	81.7% ✎	▂▄▂▂▄▇▃▇█▅▃▁ beats 92/246	very easy · avg 85%	▇▇▇ 6.5%	sharp
Decision Index 0.4.1: NLI4CT (macro-F1) submitted	74.5% ✎	▁▁▁▂▂▃▃▄▇▇█▄ beats 96/246	easy · avg 74%	▇▇ 5.1%	at its level
Decision Index 0.4.1: ANLI (macro-F1) submitted	55.7% ✎	▁▂▃▅▅█▂····▁ beats 85/246	medium · avg 60%	▇▇ 4.7%	at its level
Decision Index 0.4.1: iSarcasmEval (Sarcasm F1 · track A, English) submitted	59.0% ✎	▁·▁·▁▁▂▄▅█▇▄ beats 177/246	medium · avg 44%	▇ 1.9%	at its level
Decision Index 0.4.1: ContractNLI (macro-F1) submitted	56.3% ✎	▁▁▁▁▁▁▁▃▂▇▇█ beats 44/246	easy · avg 71%	▇ 1.0%	at its level · noisy
Decision Index 0.4.1: VAST (macro-F1) submitted	60.9% ✎	▁▁▁▂▅▄▄▃▄▆█▄ beats 110/246	medium · avg 62%	▇ 0.9%	at its level
Decision Index 0.4.1: ACOS (per-review F1) submitted	14.4% ✎	▁··▂▃▃▅▇█▇▃▃ beats 78/246	hard · avg 22%	▇ 0.8%	noisy
Decision Index 0.4.1: RAGTruth (F1 on hallucinated class) submitted	67.9% ✎	▁▁▁▁▁▁▁▁▁▃▆█ beats 89/246	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	89.8% ✎	▁···▁▁▁▁▃▆█▃ beats 144/246	easy · avg 84%	▇ 0.7%	noisy

arts & human taste▁▇▄▃▅▅█▅▄▂▁▁ #123 of 247 · 7.6% of its score · 5 tests

Beats Bobcat 1.3 (31B), JEV-35B, OneJev1.6-4B here, though they rank higher overall. Trails vLLM-SR vela3-fast-vision (Vela 3.0, 0.25B), Crivo-400M, Amber 0.4B 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	57.4% ✎	▁▁▂▃▃▃▇▅█▅▃▁ beats 83/246	medium · avg 61%	▇ 3.6%	at its level
Decision Index 0.4.1: Humicroedit (accuracy) submitted	60.8% ✎	▁·▁▃▂▂▅█▄▂▁▁ beats 114/246	medium · avg 60%	▇ 1.6%	blunt · at its level
Decision Index 0.4.1: BPoMP (accuracy) submitted	88.5% ✎	▁█▄▄▅█▆▅▄▁▁▁ beats 124/246	easy · avg 83%	▇ 1.3%	sharp · noisy
Decision Index 0.4.1: cfcolor (accuracy) submitted	61.0% ✎	▁········▁▇█ beats 112/246	medium · avg 60%	▇ 0.9%	blunt · at its level
Decision Index 0.4.1: Habermas Machine (accuracy) submitted	46.7% ✎	▁▂▄█▇▅▁▁▄▄▁▁ beats 156/246	medium · avg 47%	▇ 0.3%	blunt · at its level

knowledge & reasoning▁▂▂▂▂▃▃▅▅▆█▂ #136 of 263 · 43% of its score · 14 tests

Beats Winnow-12B (DecisionBench b6ac22b0), Amber Hybrid 4B (decision readout), StartLux-Decision v1.1 2B here, though they rank higher overall.

test	result	vs the others in its run	difficulty	counts	why
Decision Index 0.4.1: MMLU-Pro (accuracy) submitted	41.3% ✎	▄▄▃▃▂▄▅▇██▅▁ beats 84/246	medium · avg 46%	▇▇▇ 8.4%	sharp · at its level
Decision Index 0.4.1: MMLU (accuracy) submitted	78.6% ✎	▁▄▄▃▃▂▅▇▅█▇▁ beats 144/246	easy · avg 70%	▇▇▇ 6.9%	sharp · at its level
Decision Index 0.4.1: BBH (accuracy) submitted	66.4% ✎	▄▆▄▄█▇▇▄▃▁▁▁ beats 111/246	easy · avg 66%	▇▇ 5.9%	at its level
Decision Index 0.4.1: ARC-Challenge (accuracy) submitted	94.0% ✎	▂▂▂▂▂▂▁▃▄▆█▆ beats 126/246	very easy · avg 90%	▇▇ 5.5%	sharp
Decision Index 0.4.1: ARC-Easy (accuracy) submitted	98.1% ✎	▁▁▂▁▂▂▂▂▃▆█▅ beats 132/246	very easy · avg 96%	▇ 3.3%	sharp · maxed out · noisy
Decision Index 0.4.1: WinoGrande (accuracy) submitted	70.8% ✎	█▅█▅▄█▆▄▃▃▁▁ beats 91/246	easy · avg 78%	▇ 3.1%	typical
Decision Index 0.4.1: GSM8K (accuracy (4- and 10-choice)) submitted	20.7% ✎	▁··▄█▆▆▆▅▅▃▁ beats 43/246	medium · avg 45%	▇ 2.8%	sharp · at its level
Decision Index 0.4.1: GPQA Diamond (accuracy) submitted	49.0% ✎	▂▆▅▅▆█▆▃▁··▁ beats 189/246	medium · avg 39%	▇ 2.5%	blunt · at its level

6 more tests

test	result	vs the others in its run	difficulty	counts	why
Decision Index 0.4.1: MuSR (accuracy) submitted	62.5% ✎	▁▁▂█▅▄▁▁▁▁▁▁ beats 154/246	medium · avg 60%	▇ 1.0%	blunt · at its level
Decision Index 0.4.1: SATA-Bench (case exact accuracy) submitted	25.6% ✎	▁▁▁▁▁▁▁▂▅█▂▁ beats 108/246	hard · avg 22%	▇ 1.0%	typical
Decision Index 0.4.1: CRUXEval (accuracy) submitted	45.3% ✎	██▄▃▃▄▁▁···▁ beats 90/246	medium · avg 58%	▇ 1.0%	at its level · noisy
Decision Index 0.4.1: CLadder (accuracy) submitted	63.6% ✎	▁███▆▅▃▃▁▁▁▁ beats 85/246	easy · avg 71%	▇ 0.9%	at its level
Decision Index 0.4.1: ChessBench (accuracy) submitted	21.7% ✎	▄▇▄▄▄▆▄█▄▁▁▁ beats 182/246	hard · avg 16%	▇ 0.5%	blunt
Decision Index 0.4.1: HLE (accuracy) submitted	10.4% ✎	▁▁▂▃▆█▇▅▅▃▁▁ beats 43/246	very hard · avg 13%	▇ 0.0%	blunt

vision▁·▃▃▃▃▇▄▄█▄▃ #69 of 79 · 10% of its score · 12 tests

Beats StartLux-Decision v1.1 0.8B here, though they rank higher overall. Trails LiquidAI d1-omni-600M, InternLM Intern-Decision Qwen3.5-2B, Valen-2B 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	68.4800 points ✎	▁·▁▁▃▁▄▄▃▅█▂ beats 24/61	avg 63.423 points	▇ 1.3%	blunt
Decision Index Vision 0.4 · MMMU-Pro vision (chance-corrected points) submitted	25.2200 points ✎	▁▁▅▁▄▁▄█▄▃▇▇ beats 19/61	avg 22.797 points	▇ 1.2%	blunt
Decision Index Vision 0.4 · Winoground (chance-corrected points) submitted	41.7500 points ✎	▁·▃▁▁▁·▂▁▃▅█ beats 1/61	avg 66.987 points	▇ 1.1%	blunt
Decision Index Vision 0.4: private same skills (points)	39.2300 points ✎	▁▁▃▃▁▂▄▆▃▃▅█ beats 7/61	avg 48.067 points	▇ 1.1%	blunt
Decision Index Vision 0.4 · Moderation (Hateful Memes) (chance-corrected points) submitted	36.1900 points ✎	▆▃▃▆▄▄▇██▁▆▆ beats 28/61	avg 28.443 points	▇ 0.9%	blunt
Decision Index Vision 0.4 · RealWorldQA (chance-corrected points) submitted	46.4500 points ✎	▁·▂▂▃▃▃▅▅▃▂█ beats 8/61	avg 53.651 points	▇ 0.9%	blunt
Decision Index Vision 0.4 · CharXiv (chance-corrected points) submitted	32.8300 points ✎	▃▁▁▁▁▃▁▅▅▆██ beats 0/61	avg 62.137 points	▇ 0.8%	blunt
Decision Index Vision 0.4 · InfographicVQA (chance-corrected points) submitted	71.0400 points ✎	▁·▃▂▁▁▃▁▂▂▅█ beats 2/61	avg 83.859 points	▇ 0.7%	blunt

4 more tests

test	result	vs the others in its run	difficulty	counts	why
Decision Index Vision 0.4 · R-Bench-M (chance-corrected points) submitted	23.3000 points ✎	▂▄▇▆█▄▄▇▂··▁ beats 28/61	avg 19.676 points	▇ 0.7%	blunt
Decision Index Vision 0.4 · BLINK (chance-corrected points) submitted	53.4400 points ✎	▁···▁▂▂▂█▄▂▃ beats 30/61	avg 44.994 points	▇ 0.6%	blunt
Decision Index Vision 0.4 · CV-Bench (chance-corrected points) submitted	64.0600 points ✎	▁▁▁▁▂▃▁▁▄█▃▃ beats 10/61	avg 67.662 points	▇ 0.5%	blunt
Decision Index Vision 0.4 · KIE (CORD+FUNSD) (chance-corrected points) submitted	88.9100 points ✎	▁··▁▁▁▁▁▂▄█▁ beats 16/61	avg 85.184 points	▇ 0.4%	blunt

other tests2 not in a task · 6.8% of its score

test	result	vs the others in its run	difficulty	counts	why
Decision Index 0.4.1: private same skills (equated points)	32.7800 points ✎	▂▂▂▃▃▃▄▅▅▇█▃ beats 71/247	avg 41.206 points	▇▇ 4.4%	blunt
Decision Index 0.4.1: private new domains (equated points)	36.4700 points ✎	▃█▃▃▂▆▃▃▄▄▄▄ beats 108/247	avg 38.861 points	▇ 2.4%	blunt

$ where to run it

Implementation and setup instructions ↗
Published implementation and setup documentation verified. This entry describes a decision readout or evaluated configuration, not necessarily separate trained weights. Follow the referenced version and model settings; hardware compatibility has not been verified. Source · checked 2026-10-31.

Model repository and documentation: github.com/vllm-project ↗

$ how the score was worked out · Decision Index 0.4.1 89%, Decision Index Vision 0.4 10%

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	38	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 89%	CC BY 4.0
Decision Index Vision 0.4 · submitted; independent execution unverified	12	▇▇▇ 10%	CC BY 4.0
prior · pull toward the middle		▇ 1%
weighted average		= 48.4

not counted (2)

Decision Index 0.4.1	Decision Index 0.4.1: RouterBench (selected quality (quality objective)) [correct]	79.9%	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]	43.0%	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