whatmodel.sh · find the best decision models you can run

Best decision models to run locally on a 32GB graphics card

A leaderboard of the best decision models you can run locally on a 32GB graphics card (with 32GB of system RAM). Perplexity Decider Flash v1.2 (4B) is #1 right now, followed by vLLM-SR d3-flash (9B) and Jebadiah 27B v2.1 (Ollaya Q4_K_M). Models are ranked by one score that combines 238 published benchmarks. Each model shows the download that fits and the command to run it.

28 models · 238 benchmarks · updated

Compare tested decision models and inference recipes across text, vision and workflow tasks. Use task filters and raw results: probability consistency measures coherence, not correctness.

Filter by your hardware and budget → ~32GB fast · local only · reset
ram[2][4][8][16][24][32][48][64][96][128][192][256]
[]
GB
gpu[apple silicon][graphics card][unified memory][none]
vram[4][8][12][16][24][32][48][64][80][96][128]
[]
GB
price[local][$0.05][$0.1][any]max per M input tokens
Filter by task
task[all tasks][tools & automation][retrieval & classification][language understanding][arts & human taste][knowledge & reasoning][vision][workflows][consistency]
New #1s for your hardware:
orRSS

detecting your OS…

$ terminal

detecting your OS…

Run curl "whatmodel.sh/decision?ram=32&gpu=card&vram=32&max=0"to get this list in your terminal (add &format=json for agents).

Score vs release date

Release dates available for 9 of 28 models. Undated models are omitted; month-only dates are plotted at the start of the month.

#modelscore↓where↓RAM↓speed↓get itsources
#modelscore↓where↓RAM↓speed↓get itsources
90Perplexity Decider Flash v1.2 (4B) · ollaya-f32 Perplexity62±1local~10GBhow to run › Decision Index 0.4.1System One Bench (Ali Alhashimi)Etsabary · decision benchmarksOllaya · original runsDecision Index Vision 0.4Do We Need an LLM to Know Why Sources Conflict? Decision Models for Knowledge-Conflict Typing in RAG
103vLLM-SR d3-flash (9B) · ollaya-f32 vLLM Semantic Router61±1local~18GBhow to run › Decision Index 0.4.1Ollaya · original runsDecision Index Vision 0.4
125Jebadiah 27B v2.1 (Ollaya Q4_K_M) · q4_k_m-ollaya frontier-infra59±4local~16GBhow to run › Ollaya · original runs
126vLLM-SR d3-mini (4B) · ollaya-f32 vLLM Semantic Router58±1local~11GBhow to run › Decision Index 0.4.1Ollaya · original runsDecision Index Vision 0.4
131Winnow-12B (DecisionBench b6ac22b0) · ollaya-q8-0 EldanRing58±2local~16GBhow to run › Hanno Labs DecisionBench v1 reviewed resultsOllaya · original runs
134Clef-flash · bf16 Cloudflare58±1local~25GBhow to run › DecisionPivot · StartLux LabsDecision Index 0.4.1LeanMCP · JevBenchSystem One Bench (Ali Alhashimi)Etsabary · decision benchmarksDecideBench (Cho Yin Yong)+6
135Bespoke Nimble 9B v3 · ollaya-f32 Bespoke Labs58±1local~21GBhow to run › Decision Index 0.4.1Ollaya · original runsBenchmark Heaven (own decision evaluations)
143Jebadiah 9B v2.1 (T1e) · bf16-t1e frontier-infra57±1local~25GBhow to run › Decision Index 0.4.1
160Jebadiah 9B v2.1 (Ollaya Q8_0) · q8_0-ollaya frontier-infra55±3local~9.1GBhow to run › Ollaya · original runs
176vLLM-SR Decision 2.0 Lux 9B · ollaya-f32 vLLM Semantic Router54±1local~20GBhow to run › Decision Index 0.4.1Ollaya · original runs
192Jebadiah 4B v2 (Ollaya Q8_0) · q8_0-ollaya frontier-infra52±3local~5.5GBhow to run › Ollaya · original runs
200vLLM-SR Decision 2.0 Nox 4B · ollaya-f32 vLLM Semantic Router51±1local~12GBhow to run › Decision Index 0.4.1Ollaya · original runs
201Decider 4B · ollaya-f32 Mapika51±1local~11GBhow to run › Decision Index 0.4.1Ollaya · original runsCitead GEO Citation Benchmark (October 2026)
205Bespoke Nimble 9B v2 · ollaya-f32 Bespoke Labs50±1local~20GBhow to run › Decision Index 0.4.1Ollaya · original runs
216vLLM-SR d3-nano (2B) · ollaya-f32 vLLM Semantic Router49±1local~8.2GBhow to run › Decision Index 0.4.1System One Bench (Ali Alhashimi)Ollaya · original runsDecision Index Vision 0.4
219Kev 4B r10 · ollaya-f32 Jared Palmer49±1local~11GBhow to run › Decision Index 0.4.1System One Bench (Ali Alhashimi)Ollaya · original runsJevalKit (Valen Team)
247GLiClass v4 edge (0.4B) · ollaya-f32 Knowledgator Engineering45±1local~2.7GBhow to run › Decision Index 0.4.1Ollaya · original runs
254vLLM-SR d3-lite (0.8B) · ollaya-f32 vLLM Semantic Router42±1local~4.5GBhow to run › Decision Index 0.4.1System One Bench (Ali Alhashimi)Ollaya · original runsDecision Index Vision 0.4
263Decision-1.0-Eos-0.8B (DecisionBench 3c2d6326) · ollaya-f32 llm-semantic-router41±1local~5.8GBhow to run › Hanno Labs DecisionBench v1 reviewed resultsOllaya · original runs
272vLLM-SR Decision 2.0 Sol 2B · ollaya-f32 vLLM Semantic Router38±1local~11GBhow to run › Decision Index 0.4.1Ollaya · original runs
283Strands Decider 2B (Hobson v21) · bf16 Amazon Web Services36±1local~6.6GBhow to run › Decision Index 0.4.1
294vLLM-SR d3-edge (0.6B) · ollaya-f32 vLLM Semantic Router33±1local~2.6GBhow to run › Decision Index 0.4.1System One Bench (Ali Alhashimi)Ollaya · original runsDecision Index Vision 0.4
304vLLM-SR Decision 2.0 Eos 0.8B · ollaya-f32 vLLM Semantic Router31±1local~5.8GBhow to run › Decision Index 0.4.1Ollaya · original runs
308Kev 0.8B r15 · ollaya-f32 Jared Palmer30±1local~4.6GBhow to run › Decision Index 0.4.1System One Bench (Ali Alhashimi)Ollaya · original runs
314LiquidAI d1-omni-600M · ollaya-f32 Liquid AI30±1local~2.7GBhow to run › Decision Index 0.4.1System One Bench (Ali Alhashimi)Ollaya · original runsBenchmark Heaven (own decision evaluations)
317vLLM-SR Decision 2.0 Kai 0.6B · ollaya-f32 vLLM Semantic Router29±1local~3.7GBhow to run › Decision Index 0.4.1Ollaya · original runs
331laya (DecisionBench 1c5edc17) · ollaya-f16 ConvAI Innovations25±1local~1.7GBhow to run › Hanno Labs DecisionBench v1 reviewed resultsOllaya · original runs
343CLM-v0.1-8B Qwen3-8B heads · ollaya-f32 Contrastive-LM20±1local~15GBhow to run › Decision Index 0.4.1DecideBench (Cho Yin Yong)Ollaya · original runsAgentDecisionBenchPerch benchmark results

+ 329 more don't fit your machine or budget · show them

357 of 357 catalogue models listed here · sources and licences

* new or thinly tested: ranked on too few benchmark results · † no independently run results verified · ~RAM is estimated; full context may need more memory

28 models · 238 benchmarks · updated