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

Best AI models for math to run locally on a 32GB Mac

A leaderboard of the best AI models for math you can run locally on a Mac with 32GB of memory. Granite 4.2 30B is #1 right now, followed by Qwen3.6 27B and Spark-X2.5 4B. Models are ranked on math tests, from competition problems like AIME to research problems like FrontierMath. Each model shows the download that fits and the command to run it.

213 models · scores combine 59 benchmarks · updated how it works · find a leaderboard · benchmark your machine
task[all tasks][agent tasks][coding][chat & writing][math][vision & documents][web search]

Scores use reviewed math evaluations only. Compare scores within this task. Models without task measurements are excluded.

Models for 32GB RAM · Apple Silicon → ~21GB fast · local only · reset
+ more
chip[M1 Pro][M1 Max][M2 Pro][M2 Max][M4][M5][M6 (24/32GB)]
set[all sets][OTIS mock AIME 2024–2025 · 464][MGSM · 5][MathQA · 4][AIME 2024 · 3][DeepMind Mathematics · 3][IMO-AnswerBench · 3][AIME 2024 · 2][AIME 2025 · 2][AIME 2024 · 2][AIME 2026 · 2][AIME 2024 · 2][AIME 2025 · 2][AIME 2025 · 2][GSM8K · 2][GSM8K · 2][GSM8K · 2][AIME 2024 · 2][GSM8K · 2][GSM8K · 2]

Filter the model list while keeping task scores. Models without a shared score scale are not yet calibrated; select a set to see its own scores.

Detect your exact chip or graphics card:

curl -G whatmodel.sh/ -d max=0 --data-urlencode "mac=$(sysctl -n hw.memsize machdep.cpu.brand_string)"Mac

curl -G whatmodel.sh/ -d max=0 --data-urlencode "ram=$(awk '/MemT/{print $2}' /proc/meminfo)" --data-urlencode "nv=$(nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>/dev/null)"Linux

curl.exe -G whatmodel.sh/ -d max=0 --data-urlencode "ram=$((gcim Win32_ComputerSystem).TotalPhysicalMemory)" --data-urlencode "nv=$(nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>$null)"Windows

$ terminal

detecting your OS…

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

ram[8][16][24][32][48][64][96][128][192][256]GB
gpu[apple silicon][graphics card][none]
price[local][$0.02][$0.1][$0.5][$2][$10][$50][any]max per M

detecting your OS…

Score vs memory to run

#modeltask score↓where↓speed↓RAM↓context↓get itsources
#modeltask score↓where↓speed↓RAM↓context↓get itsources
74Granite 4.2 30B · 4-bit IBM65±3*†local~21GB131kOllama › ibm-granite/granite-4.2-30b model card
84Qwen3.6 27B · 4-bit Alibaba64±1local~20GB262kOllama › Epoch AI — internal evaluationsQwen/Qwen3.6-27B model card
92Spark-X2.5 4B · 16-bit iFLYTEK (SparkLLM)64±2†local~10GB1MOllama › XHToken/Spark-X2.5-4B model card
99Qwen3.5 27B · 4-bit Alibaba63±1local~20GB262kOllama › MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasetsQwen/Qwen3.5-122B-A10B model cardQwen/Qwen3.6-27B model cardQwen/Qwen3.6-35B-A3B model cardQwen/Qwen3.5-27B model card
101Nemotron Cascade 2 30B-A3B · 4-bit NVIDIA63±2†local~20GB256kLM Studio › NVIDIA — Nemotron-Cascade 2 technical report
104Gemma 4 26B-A4B · 4-bit Google63±2local~19GB262kllama.cpp › Epoch AI — internal evaluationsgoogle/diffusiongemma-26B-A4B-it model cardGemma Team — Gemma 4 Technical Report
120Olmo 3.1 32B Think · 4-bit Ai262±2†local~21GB66kLM Studio › Olmo 3allenai/Olmo-3.1-32B-Think model card
123Gemma 4 31B · 4-bit Google61±2local~21GB262kOllama › Epoch AI — internal evaluationsGemma Team — Gemma 4 Technical Report
128Nemotron-Cascade-14B-Thinking · 8-bit NVIDIA61±2*†local~18GBllama.cpp › NVIDIA — Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models
130Granite 4.2 8B · 8-bit IBM61±3*†local~11GB131kllama.cpp › ibm-granite/granite-4.2-30b model card
131Falcon-H1R 7B · 8-bit TII61±2†local~12GB262kllama.cpp › Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time Scaling
133Qwen3-VL 30B-A3B Thinking · 4-bit Alibaba61±3*†local~21GB256kllama.cpp › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
139Nemotron-Cascade-8B · 8-bit NVIDIA60±2*†local~11GBllama.cpp › NVIDIA — Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models
143Olmo 3 32B Think · 4-bit Ai260±1†local~21GB66kLM Studio › Olmo 3allenai/Olmo-3.1-32B-Think model card
146gpt-oss-20b · MXFP4 OpenAI60±1local~15GB131kOllama › Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time ScalingByteDance-Seed/Seed-OSS-36B-Instruct model cardEpoch AI — internal evaluationsOpenAI — gpt-oss-120b & gpt-oss-20b Model Card
147MiniCPM5 2B · 4-bit OpenBMB60±2†local~4.3GB131kLM Studio › openbmb/MiniCPM5-2B model card
150Qwen3 30B-A3B Instruct 2507 · 4-bit Alibaba59±1local~21GB262kOllama › Qwen/Qwen3-30B-A3B-Instruct-2507 model cardEpoch AI — internal evaluationsQwen/Qwen3-Next-80B-A3B-Instruct model card
153Nemotron 3 Nano 30B-A3B · 4-bit NVIDIA59±2†local~20GB262kLM Studio › NVIDIA — Nemotron-Cascade 2 technical report
155Qwen3.5 9B · 4-bit Alibaba59±1local~12GB262kLM Studio › XHToken/Spark-X2.5-4B model cardMathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasetsQwen/Qwen3.5-9B model cardEpoch AI — internal evaluationsLiquidAI/LFM2.5-2.6B model cardQwen/Qwen3.5-4B model card
156Qwen3 14B · 8-bit Alibaba59±2*local~21GB33kOllama › Epoch AI — internal evaluations
157QwQ-32B · 4-bit Alibaba58±2local~21GB131kLM Studio › Epoch AI — internal evaluations2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct models
158Qwen3-VL 8B Thinking · 4-bit Alibaba58±2local~12GB256kLM Studio › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)allenai/Olmo-3-7B-Think model card
160Qwen3.5 4B · 4-bit Alibaba58±1local~6.9GB262kLM Studio › openbmb/MiniCPM5-2B model cardXHToken/Spark-X2.5-4B model cardQwen/Qwen3.5-9B model cardEpoch AI — internal evaluationsLiquidAI/LFM2.5-2.6B model cardQwen/Qwen3.5-4B model card
162EXAONE 4.0 32B · 4-bit LG AI Research58±2†local~20GB131kLM Studio › LG AI Research vendor-reported benchmark scores for its own EXAONE models (Hugging Face model cards)
163Granite 4.2 3B · 8-bit IBM58±1local~5.6GB131kllama.cpp › openbmb/MiniCPM5-2B model cardibm-granite/granite-4.2-30b model card
164DeepSeek R1 0528 Qwen3 8B · 8-bit DeepSeek58±1local~12GB128kOllama › Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time ScalingEpoch AI — internal evaluationsdeepseek-ai/DeepSeek-R1-0528-Qwen3-8B model card
165Qwen3-VL 4B Thinking · 8-bit Alibaba57±2*†local~6.9GB256kOllama › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
166Qwen3 30B-A3B Thinking 2507 · 4-bit Alibaba57±1local~21GB262kOllama › ByteDance-Seed/Seed-OSS-36B-Instruct model cardMathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasetsEpoch AI — internal evaluationsQwen/Qwen3-30B-A3B-Thinking-2507 model cardQwen/Qwen3-Next-80B-A3B-Thinking model cardQwen/Qwen3.5-9B model card
169Olmo 3 7B Think · 8-bit Ai257±1†local~11GB66kOllama › Olmo 3allenai/Olmo-3-7B-Think model card
171Phi-4-reasoning-plus 14B · 8-bit Microsoft57±1local~18GBllama.cpp › Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time ScalingMicrosoft — Phi-4-reasoning Technical Reportmicrosoft/Phi-4-reasoning model card
173GLM-4.7-Flash · 4-bit Z.ai57±2*local~21GB200kllama.cpp › Epoch AI — internal evaluations
174Gemma 4 12B · 8-bit Google57±2local~15GB262kllama.cpp › XHToken/Spark-X2.5-4B model cardGemma Team — Gemma 4 Technical Report
177Nemotron Nano 9B v2 · 8-bit NVIDIA57±1local~13GB131kllama.cpp › Olmo 3NVIDIA — Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Modelsallenai/Olmo-3-7B-Think model cardNVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards)
179Qwen3-VL 30B-A3B Instruct · 4-bit Alibaba56±2*†local~21GB256kllama.cpp › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
184OpenThinker3 7B · 8-bit Open Thoughts56±1local~10GB33kllama.cpp › Olmo 3allenai/Olmo-3-7B-Think model card
185OpenThinker2 32B · 4-bit Open Thoughts56±2†local~21GB33kLM Studio › open-thoughts/OpenThinker2-32B model card
186Qwen3 32B · 4-bit Alibaba56±1local~20GB131kLM Studio › Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time ScalingByteDance-Seed/Seed-OSS-36B-Instruct model cardEpoch AI — internal evaluationsQwen/Qwen3-Next-80B-A3B-Thinking model cardallenai/Olmo-3.1-32B-Think model cardQwen/Qwen3-Next-80B-A3B-Instruct model card
187Qwen3 8B · 8-bit Alibaba55±1local~12GB33kOllama › Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time Scalingallenai/Olmo-3-7B-Think model cardEpoch AI — internal evaluationsQwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)allenai/Olmo-3-7B-Instruct model card
188DeepSeek R1 Distill Qwen 32B · 4-bit DeepSeek55±1local~21GB131kLM Studio › deepseek-ai/DeepSeek-R1 model cardopen-thoughts/OpenThinker2-32B model cardEpoch AI — internal evaluationsallenai/Olmo-3.1-32B-Think model cardopen-thoughts/OpenThinker-32B model cardTHINK-Bench: Evaluating Thinking Efficiency and Chain-of-Thought Quality of Large Reasoning Models (arXiv:2505.22113v1)
189DeepSeek R1 Distill Qwen 14B · 8-bit DeepSeek55±1local~21GB131kOllama › Epoch AI — internal evaluationsdeepseek-ai/DeepSeek-R1 model cardTHINK-Bench: Evaluating Thinking Efficiency and Chain-of-Thought Quality of Large Reasoning Models (arXiv:2505.22113v1)
190Phi-4-reasoning 14B · 8-bit Microsoft55±2†local~21GB33kOllama › Microsoft — Phi-4-reasoning Technical Reportmicrosoft/Phi-4-reasoning model card
196Qwen3 4B Instruct 2507 · 8-bit Alibaba54±1local~7.1GB262kOllama › Marco-MoE: Open Multilingual Mixture-of-Expert Language Models with Efficient UpcyclingEpoch AI — internal evaluationsQwen/Qwen3-4B-Instruct-2507 model cardQwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
197Phi-4-mini-reasoning 3.8B · 4-bit Microsoft54±2*†local~5.8GB128kLM Studio › microsoft/Phi-4-mini-reasoning model card
198Magistral Small 1.2 · 4-bit Mistral54±1local~17GB131kllama.cpp › mistralai/Magistral-Small-2509 model cardEpoch AI — internal evaluations
201OpenThinker 32B · 4-bit Open Thoughts53±1†local~21GB33kLM Studio › open-thoughts/OpenThinker2-32B model cardopen-thoughts/OpenThinker-32B model card
202Qwen3 4B Thinking 2507 · 8-bit Alibaba53±2local~7.1GB262kOllama › Qwen/Qwen3-4B-Thinking-2507 model cardMathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets
206QED-Nano · 4-bit LM Provers52±2local~6GB262kLM Studio › MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets
207Qwen3-VL 8B Instruct · 4-bit Alibaba52±2local~12GB256kLM Studio › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)allenai/Olmo-3-7B-Instruct model card
210DeepSeek R1 Distill Qwen 7B · 8-bit DeepSeek52±1local~12GB131kOllama › deepseek-ai/DeepSeek-R1 model cardopen-thoughts/OpenThinker2-7B model cardallenai/Olmo-3-7B-Think model cardmicrosoft/Phi-4-mini-reasoning model cardMicrosoft — Phi-4-Mini Technical Reportopen-thoughts/OpenThinker-7B model card+1
212Nemotron 3 Nano 4B · 8-bit NVIDIA51±1local~7GB262kOllama › openbmb/MiniCPM5-2B model cardNVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards)
214Spark-X2.5 1.7B · 16-bit iFLYTEK (SparkLLM)51±2†local~5.1GB1MOllama › XHToken/Spark-X2.5-4B model card
217Qwen3 30B-A3B (Apr 2025, hybrid) · 4-bit Alibaba50±1local~21GB33kOllama › Qwen/Qwen3-4B-Thinking-2507 model cardQwen/Qwen3-30B-A3B-Thinking-2507 model cardQwen/Qwen3-30B-A3B-Instruct-2507 model cardQwen/Qwen3-4B-Instruct-2507 model cardEpoch AI — internal evaluations
219EXAONE Deep 7.8B · 4-bit LG AI Research50±2†local~10GB33kLM Studio › LG AI Research vendor-reported benchmark scores for its own EXAONE models (Hugging Face model cards)
220OpenThinker2 7B · 4-bit Open Thoughts50±2†local~6.4GB33kllama.cpp › open-thoughts/OpenThinker2-7B model card
222Gemma 3 27B · 4-bit Google50±1local~19GB131kOllama › Epoch AI — internal evaluationsallenai/Olmo-3.1-32B-Think model cardGoogle DeepMind Gemma 3 270M IT model card2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct modelsGemma Team — Gemma 4 Technical Report
223Ministral 3 14B · 8-bit Mistral50±3*†local~20GB262kOllama › mistralai/Ministral-3-8B-Instruct-2512 model cardmistralai/Ministral-3-14B-Instruct-2512 model card
224Magistral Small 1.0 (2506) · 4-bit Mistral50±1local~17GB128kOllama › mistralai/Magistral-Small-2509 model cardEpoch AI — internal evaluations
225Mistral Small 3.2 24B · 4-bit Mistral50±2*local~17GB128kOllama › Epoch AI — internal evaluationsmistralai/Mistral-Small-3.2-24B-Instruct-2506 model card
228Olmo 3 7B Instruct · 8-bit Ai249±1†local~11GB66kOllama › Olmo 3allenai/Olmo-3-7B-Instruct model card
231Gemma 3 12B · 8-bit Google49±1local~17GB131kOllama › Marco-MoE: Open Multilingual Mixture-of-Expert Language Models with Efficient UpcyclingEpoch AI — internal evaluationsGoogle DeepMind Gemma 3 270M IT model card
232DeepSeek R1 Distill Llama 8B · 8-bit DeepSeek49±1local~12GB131kOllama › deepseek-ai/DeepSeek-R1 model cardmicrosoft/Phi-4-mini-reasoning model cardMicrosoft — Phi-4-Mini Technical Report
233Magistral Small 1.1 (2507) · 4-bit Mistral49±2*†local~17GB40kllama.cpp › mistralai/Magistral-Small-2509 model card
235Qwen3-VL 4B Instruct · 8-bit Alibaba49±2*†local~6.9GB256kOllama › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
239LFM2.5 1.2B Thinking · 8-bit Liquid AI49±3†local~3.5GB33kOllama › LiquidAI/LFM2.5-1.2B-Thinking model card
240Ministral 3 8B · 8-bit Mistral48±2local~13GB262kOllama › Marco-MoE: Open Multilingual Mixture-of-Expert Language Models with Efficient Upcyclingmistralai/Ministral-3-8B-Instruct-2512 model cardmistralai/Ministral-3-14B-Instruct-2512 model card
242LFM2.5 8B-A1B · 8-bit Liquid AI48±1local~13GB128kOllama › openbmb/MiniCPM5-2B model cardLiquidAI/LFM2.5-8B-A1B model card
244Qwen3 4B · 8-bit Alibaba48±1local~7.1GB33kOllama › Qwen/Qwen3-4B-Thinking-2507 model cardQwen/Qwen3-4B-Instruct-2507 model cardLFM2 Technical Report — Liquid AI own SmolLM3 evaluationHuggingFaceTB/SmolLM3-3B model card (Evaluation section)
245LFM2 24B-A2B · 4-bit Liquid AI48±2local~17GB33kOllama › Marco-MoE: Open Multilingual Mixture-of-Expert Language Models with Efficient Upcycling
249Phi-4 14B · 8-bit Microsoft48±1local~21GB16kOllama › Epoch AI — internal evaluationsmicrosoft/phi-4 model cardMicrosoft Research — Phi-4 Technical Report
251Qwen3-VL 2B Thinking · 8-bit Alibaba48±2*†local~4.2GB256kOllama › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
258Granite 4.0 H Small (32B-A9B) · 4-bit IBM47±1local~21GB131kLM Studio › Marco-MoE: Open Multilingual Mixture-of-Expert Language Models with Efficient Upcyclingibm-granite/granite-4.0-h-small model card
262OpenThinker 7B · 8-bit Open Thoughts46±2†local~10GB33kllama.cpp › open-thoughts/OpenThinker2-7B model cardopen-thoughts/OpenThinker-7B model card
264EXAONE Deep 2.4B · 4-bit LG AI Research46±2†local~3.2GB33kllama.cpp › LG AI Research vendor-reported benchmark scores for its own EXAONE models (Hugging Face model cards)
265Qwen2.5-Math 7B Instruct · 4-bit Alibaba46±2†local~10GB4kLM Studio › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
266Qwen2.5 32B Instruct · 4-bit Alibaba46±1local~21GB131kLM Studio › Epoch AI — internal evaluationsWildEval ZeroEval published resultsallenai/Olmo-3.1-32B-Think model card2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct models
268LFM2.5 2.6B · 4-bit Liquid AI46±1local~4.5GB131kLM Studio › openbmb/MiniCPM5-2B model cardLiquidAI/LFM2.5-2.6B model card
273Qwen2.5 14B Instruct · 8-bit Alibaba45±2local~21GB131kOllama › 2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct modelsmicrosoft/phi-4 model cardMicrosoft Research — Phi-4 Technical Report
274Granite 4.0 H Tiny (7B-A1B) · 4-bit IBM45±2†local~9.3GB131kLM Studio › ibm-granite/granite-4.0-h-small model card
275DeepSeek R1 Distill Qwen 1.5B · 8-bit DeepSeek45±1local~4.3GB131kOllama › deepseek-ai/DeepSeek-R1 model cardEpoch AI — internal evaluationsTHINK-Bench: Evaluating Thinking Efficiency and Chain-of-Thought Quality of Large Reasoning Models (arXiv:2505.22113v1)
282Qwen2.5-VL 7B Instruct · 8-bit Alibaba45±3*†local~11GBOllama › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
284Gemma 4 E4B · 4-bit Google44±1local~11GB131kLM Studio › openbmb/MiniCPM5-2B model cardXHToken/Spark-X2.5-4B model cardGemma Team — Gemma 4 Technical ReportLiquidAI/LFM2.5-2.6B model card
291EXAONE 4.0 1.2B · 8-bit LG AI Research44±2*†local~3.6GB66kllama.cpp › LG AI Research vendor-reported benchmark scores for its own EXAONE models (Hugging Face model cards)
292Qwen2.5 7B Instruct · 8-bit Alibaba44±1local~12GB131kOllama › WildEval ZeroEval published resultsmicrosoft/Phi-4-mini-instruct model card2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct modelsallenai/Olmo-3-7B-Instruct model cardEpoch AI — internal evaluations
293SmolLM3 3B · 8-bit Hugging Face44±2local~5.9GB66kllama.cpp › LFM2 Technical Report — Liquid AI own SmolLM3 evaluationHuggingFaceTB/SmolLM3-3B model card (Evaluation section)
294Gemma 3 4B · 8-bit Google43±2local~7.4GB131kOllama › Google DeepMind Gemma 3 270M IT model cardLFM2 Technical Report — Liquid AI own SmolLM3 evaluationEpoch AI — internal evaluations
295Mistral Small 3.1 24B (2503) · 4-bit Mistral43±1local~17GB128kLM Studio › Epoch AI — internal evaluationsmistralai/Mistral-Small-3.2-24B-Instruct-2506 model cardMistral AI vendor-reported benchmark scores for its own models (Hugging Face model cards)
297Granite 3.3 8B Instruct · 8-bit IBM43±1local~11GB128kllama.cpp › ibm-granite/granite-3.3-8b-instruct model cardallenai/Olmo-3-7B-Instruct model card
305Qwen3 1.7B · 8-bit Alibaba43±1local~4.6GB33kOllama › Marco-MoE: Open Multilingual Mixture-of-Expert Language Models with Efficient UpcyclingLiquidAI/LFM2.5-1.2B-Thinking model cardQwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)LiquidAI/LFM2-1.2B model cardQwen/Qwen3.5-2B model cardHuggingFaceTB/SmolLM3-3B model card (Evaluation section)+1
306Mistral Small 3 24B (2501) · 4-bit Mistral43±1*local~17GB32kOllama › Epoch AI — internal evaluations
307Granite 4.0 H Micro 3B · 4-bit IBM43±2†local~5.1GB131kLM Studio › ibm-granite/granite-4.0-h-small model card
310Ministral 3 3B · 8-bit Mistral42±1local~6.9GB262kOllama › Marco-MoE: Open Multilingual Mixture-of-Expert Language Models with Efficient Upcyclingmistralai/Ministral-3-8B-Instruct-2512 model cardmistralai/Ministral-3-14B-Instruct-2512 model card
312Qwen2-Math 7B Instruct · 8-bit Alibaba42±2†local~10GB4kllama.cpp › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
313Granite 4.0 Micro 3B · 8-bit IBM42±2local~5.3GB131kllama.cpp › ibm-granite/granite-4.0-h-small model cardEpoch AI — internal evaluations
315Granite 8B Code Base (4K) · 4-bit IBM42±2*†local~6.6GBllama.cpp › Granite Code Models: A Family of Open Foundation Models for Code Intelligence — authors' own runs
322Gemma 4 E2B · 4-bit Google41±1local~7.7GB131kLM Studio › openbmb/MiniCPM5-2B model cardXHToken/Spark-X2.5-4B model cardGemma Team — Gemma 4 Technical ReportLiquidAI/LFM2.5-2.6B model card
323Qwen2.5-Math 1.5B Instruct · 4-bit Alibaba41±2†local~2.5GB4kllama.cpp › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
325Phi-4-mini 3.8B · 8-bit Microsoft41±2†local~6.9GB128kOllama › microsoft/Phi-4-mini-instruct model cardMicrosoft — Phi-4-Mini Technical Report
327Phi-2 (2.7B) · 4-bit Microsoft40±3*local~7.3GB2khow to run › MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies (Hu et al.)
328Molmo 2 8B · 8-bit Ai240±7*†local~11GBllama.cpp › Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding (Ai2, arXiv 2601.10611v4)
329Pixtral 12B (2409) · 8-bit Mistral40±3*†local~15GBllama.cpp › Pixtral 12B (Mistral AI, arXiv 2410.07073v2)
330StarCoder2 15B · 8-bit BigCode40±2*†local~19GB16kOllama › StarCoder 2 and The Stack v2: The Next Generation — authors' own runs
331LFM2 2.6B · 4-bit Liquid AI40±3*†local~4.3GB33kLM Studio › LFM2 Technical Report — Liquid AI own SmolLM3 evaluation
332Qwen3.5 2B · 8-bit Alibaba40±1local~4.9GB262kOllama › openbmb/MiniCPM5-2B model cardXHToken/Spark-X2.5-4B model cardQwen/Qwen3.5-2B model cardMathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets
334Qwen2.5-VL 3B Instruct · 8-bit Alibaba40±3*†local~6.3GBOllama › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
336Qwen3-VL 2B Instruct · 8-bit Alibaba40±2*†local~4.2GB256kOllama › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
337Gemma 7B (base) · 8-bit Google40±2local~11GB8kOllama › Granite Code Models: A Family of Open Foundation Models for Code Intelligence — authors' own runsGemma report — original evaluations
338InternLM2.5 7B Chat · 8-bit Shanghai AI Laboratory (InternLM)39±3*†local~12GBOllama › InternLM GitHub repository model cards (Shanghai AI Laboratory)
339Qwen2-Math 1.5B Instruct · 8-bit Alibaba39±2†local~3.2GB4kllama.cpp › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
340CodeGemma 7B (v1.0) · 8-bit Google39±2*local~11GBOllama › Granite Code Models: A Family of Open Foundation Models for Code Intelligence — authors' own runs
342LFM2 1.2B · 4-bit Liquid AI39±3*†local~2.8GB33kLM Studio › LiquidAI/LFM2-1.2B model card
345Llama 3.2 11B Vision Instruct · 8-bit Meta38±3*local~16GB128kOllama › Pixtral 12B (Mistral AI, arXiv 2410.07073v2)
347LFM2.5 1.2B Instruct · 4-bit Liquid AI38±2†local~2.8GB33kLM Studio › LiquidAI/LFM2.5-1.2B-Thinking model card
348Gemma 2 27B IT · 4-bit Google38±1local~19GB8kOllama › Epoch AI — internal evaluationsWildEval ZeroEval published resultsallenai/Olmo-3.1-32B-Think model card2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct models
352Mathstral 7B v0.1 · 8-bit Mistral AI38±3*†local~11GB32kOllama › Mistral AI vendor-reported benchmark scores for its own models (Hugging Face model cards)
353Granite 3.3 2B Instruct · 8-bit IBM38±2†local~4.3GB128kllama.cpp › ibm-granite/granite-3.3-8b-instruct model card
358Tulu 2 DPO 13B · 8-bit Allen Institute for AI37±5*†local~16GBllama.cpp › Tulu 2 — base and instruction-tuned evaluations
362Llama 3.1 Tulu 3 8B · 8-bit Ai237±1local~12GB131kOllama › microsoft/Phi-4-mini-instruct model card2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct models
364Granite 3.2 8B Instruct · 8-bit IBM37±2†local~12GB131kOllama › ibm-granite/granite-3.3-8b-instruct model card
366Granite 4.0 1B · 16-bit IBM37±2local~6.1GB128kOllama › ibm-granite/granite-4.0-h-1b model cardLiquidAI/LFM2.5-1.2B-Thinking model cardEpoch AI — internal evaluations
367Qwen2.5 3B Instruct · 8-bit Alibaba37±1local~5.9GB33kOllama › WildEval ZeroEval published resultsmicrosoft/Phi-4-mini-instruct model card
368OLMo 2 32B Instruct (0325) · 4-bit Ai237±1†local~21GBllama.cpp › allenai/Olmo-3.1-32B-Think model card2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct models
369Llama 3.1 8B Instruct · 8-bit Meta37±1local~12GB128kOllama › Epoch AI — internal evaluationsWildEval ZeroEval published resultsmicrosoft/Phi-4-mini-instruct model cardIBM Granite 3.0 Language Models technical report (GitHub)2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct models
370Granite 3.1 8B Instruct · 8-bit IBM36±2†local~12GB128kOllama › ibm-granite/granite-3.3-8b-instruct model cardibm-granite/granite-3.1-3b-a800m-instruct model card
371OLMo 2 13B Instruct (1124) · 8-bit Allen Institute for AI36±2*†local~19GB4kOllama › 2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct models
372Gemma 2 9B IT · 8-bit Google36±1local~14GB8kOllama › Epoch AI — internal evaluationsWildEval ZeroEval published resultsmicrosoft/Phi-4-mini-instruct model card2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct modelsInternLM GitHub repository model cards (Shanghai AI Laboratory)
373Tulu 2 13B · 8-bit Allen Institute for AI36±4*†local~16GBllama.cpp › Tulu 2 — base and instruction-tuned evaluations
374Granite 3.0 8B Instruct · 8-bit IBM36±2*†local~12GB4kOllama › IBM Granite 3.0 Language Models technical report (GitHub)
377Granite 4.0 H 1B · 8-bit IBM36±2local~3.2GB131kOllama › ibm-granite/granite-4.0-h-1b model cardLiquidAI/LFM2.5-1.2B-Thinking model card
379Qwen3 0.6B · 8-bit Alibaba35±2local~3GB33kOllama › Olmo 3LiquidAI/LFM2-1.2B model card
380Phi-3 Medium 128K Instruct · 8-bit Microsoft35±1*local~20GB128kOllama › Epoch AI — internal evaluations
382Qwen2 7B Instruct · 8-bit Alibaba35±1local~12GB131kOllama › WildEval ZeroEval published resultsQwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)InternLM GitHub repository model cards (Shanghai AI Laboratory)
384Ministral 8B (2410) · 8-bit Mistral34±1local~10GB128kllama.cpp › Epoch AI — internal evaluationsmicrosoft/Phi-4-mini-instruct model card2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct modelsMistral AI vendor-reported benchmark scores for its own models (Hugging Face model cards)
387LFM2 700M · 4-bit Liquid AI34±3*†local~2.3GB33kLM Studio › LiquidAI/LFM2-1.2B model card
390Phi-3.5-mini-instruct · 8-bit Microsoft34±1local~6.8GB128kOllama › WildEval ZeroEval published resultsmicrosoft/Phi-4-mini-instruct model card
392Qwen2-VL 7B Instruct · 8-bit Alibaba34±2local~11GBllama.cpp › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)Pixtral 12B (Mistral AI, arXiv 2410.07073v2)
394Yi 1.5 9B Chat · 8-bit 01.AI33±1*local~13GB4kOllama › WildEval ZeroEval published results
396Granite 3.0 2B Instruct · 8-bit IBM33±2*†local~5.3GB4kOllama › IBM Granite 3.0 Language Models technical report (GitHub)
397OLMo 2 7B Instruct (1124) · 8-bit Allen Institute for AI33±2†local~11GB4kOllama › 2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct modelsallenai/Olmo-3-7B-Instruct model card
398Qwen3.5 0.8B · 8-bit Alibaba33±4†local~3.1GB262kOllama › Qwen/Qwen3.5-2B model card
399Qwen2.5 1.5B Instruct · 8-bit Alibaba33±2local~4GB33kOllama › 2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct modelsSmolLM2 paper (Hugging Face) — authors' own instruct-model evaluations
401Mistral 7B v0.1 (base) · 16-bit Mistral AI33±1local~17GB8khow to run › Mixtral of Experts — base-model evaluationsMistral 7B — authors' own runsMiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies (Hu et al.)Gemma report — original evaluations
402Llama 3.2 3B Instruct · 8-bit Meta32±1local~6.1GB131kOllama › microsoft/Phi-4-mini-instruct model cardIBM Granite 3.0 Language Models technical report (GitHub)microsoft/Phi-4-mini-reasoning model cardMicrosoft — Phi-4-Mini Technical ReportLFM2 Technical Report — Liquid AI own SmolLM3 evaluation
404Mistral NeMo Instruct (2407) · 8-bit Mistral32±1*local~18GB128kOllama › Epoch AI — internal evaluationsWildEval ZeroEval published results
406Tulu 2 DPO 7B · 16-bit Allen Institute for AI32±5*†local~16GBhow to run › Tulu 2 — base and instruction-tuned evaluations
408Tulu 2 7B · 16-bit Allen Institute for AI32±5*†local~16GBhow to run › Tulu 2 — base and instruction-tuned evaluations
409Phi-3-mini-4k-instruct (June 2024 update) · 4-bit Microsoft32±1*local~5.8GBLM Studio › WildEval ZeroEval published results
411Code Tulu 2 7B · 16-bit Allen Institute for AI32±5*†local~16GBhow to run › Tulu 2 — base and instruction-tuned evaluations
412Gemma 3 1B · 4-bit Google31±1local~3GB32kLM Studio › Google DeepMind Gemma 3 270M IT model cardLiquidAI/LFM2.5-1.2B-Thinking model card2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct modelsLiquidAI/LFM2-1.2B model cardEpoch AI — internal evaluations
413DeepSeek-Coder 6.7B Base · 8-bit DeepSeek31±2*local~9.1GB16kOllama › StarCoder 2 and The Stack v2: The Next Generation — authors' own runs
414Granite 3.2 2B Instruct · 8-bit IBM31±2†local~5.2GB131kOllama › ibm-granite/granite-3.3-8b-instruct model card
415Granite 3.1 2B Instruct · 8-bit IBM31±2†local~5.2GB128kOllama › ibm-granite/granite-3.3-8b-instruct model cardibm-granite/granite-3.1-3b-a800m-instruct model card
417Stable Code 3B · 8-bit Stability AI31±2*local~4.7GB16kOllama › StarCoder 2 and The Stack v2: The Next Generation — authors' own runs
419StarCoder2 7B · 8-bit BigCode30±2local~9.5GB16kOllama › StarCoder 2 and The Stack v2: The Next Generation — authors' own runsGranite Code Models: A Family of Open Foundation Models for Code Intelligence — authors' own runs
420Granite 3.1 3B-A800M Instruct · 8-bit IBM30±4*†local~5.2GB131kllama.cpp › ibm-granite/granite-3.1-3b-a800m-instruct model card
421Zephyr 7B beta · 8-bit Hugging Face30±5*local~11GBOllama › Tulu 2 — base and instruction-tuned evaluations
422Granite 3.0 3B-A800M Instruct · 8-bit IBM30±2*†local~5.3GB4kllama.cpp › IBM Granite 3.0 Language Models technical report (GitHub)
423OLMo 2 1B Instruct (0425) · 8-bit Allen Institute for AI29±2*†local~3.2GB4kllama.cpp › 2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct models
425Code Llama 13B (base) · 8-bit Meta29±2*local~16GB16kOllama › StarCoder 2 and The Stack v2: The Next Generation — authors' own runsTulu 2 — base and instruction-tuned evaluations
426Llama 3 8B Instruct · 8-bit Meta29±1local~12GB8kOllama › Epoch AI — internal evaluationsWildEval ZeroEval published resultsInternLM GitHub repository model cards (Shanghai AI Laboratory)
427WizardLM 13B V1.2 · 8-bit Microsoft29±4*local~16GBllama.cpp › Orca 2: Teaching Small Language Models How to Reason (Mitra et al., Microsoft)
429Apertus 8B Instruct (2509) · 8-bit Swiss AI29±3*local~11GB66kllama.cpp › allenai/Olmo-3-7B-Instruct model card
430Code Llama 13B Instruct · 8-bit Meta28±5*local~16GB16kOllama › Tulu 2 — base and instruction-tuned evaluations
431SmolLM2 1.7B Instruct · 8-bit Hugging Face27±2local~4.2GB8kOllama › 2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct modelsSmolLM2 paper (Hugging Face) — authors' own instruct-model evaluations
433LFM2 350M · 4-bit Liquid AI27±3*†local~1.9GB33kLM Studio › LiquidAI/LFM2-1.2B model card
435Gemma 2B (base) · 8-bit Google27±4*†local~4.3GB8kOllama › Gemma report — original evaluations
437Llama 2 13B (base) · 8-bit Meta26±2local~16GB4kllama.cpp › Mixtral of Experts — base-model evaluationsMistral 7B — authors' own runsMiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies (Hu et al.)Tulu 2 — base and instruction-tuned evaluations
438StarCoder2 3B · 4-bit BigCode26±2*†local~5.1GB16kLM Studio › StarCoder 2 and The Stack v2: The Next Generation — authors' own runs
439Llama 3.2 1B Instruct · 4-bit Meta25±1local~5.2GB131kOllama › IBM Granite 3.0 Language Models technical report (GitHub)LiquidAI/LFM2.5-1.2B-Thinking model card2 OLMo 2 Furious (Team OLMo, Ai2) — authors' own evaluations of OLMo 2 Instruct and peer instruct modelsSmolLM2 paper (Hugging Face) — authors' own instruct-model evaluationsLiquidAI/LFM2-1.2B model cardEpoch AI — internal evaluations
440Code Llama 7B Instruct · 8-bit Meta25±5*local~9.1GB16kOllama › Tulu 2 — base and instruction-tuned evaluations
441Llama-3-Instruct-8B-SimPO-v0.2 · 16-bit Princeton NLP25±2*local~18GBhow to run › WildEval ZeroEval published results
442Granite 3.1 1B-A400M Instruct · 8-bit IBM25±4*†local~3GB131kOllama › ibm-granite/granite-3.1-3b-a800m-instruct model card
443Mistral 7B Instruct v0.3 · 8-bit Mistral25±1local~11GB32kOllama › Epoch AI — internal evaluationsMistral AI vendor-reported benchmark scores for its own models (Hugging Face model cards)
444Qwen2 0.5B Instruct · 8-bit Alibaba24±5*†local~2.6GB32kOllama › Qwen2-0.5B-Instruct model card
446Gemma 2 2B IT · 8-bit Google24±1local~5.3GB8kOllama › WildEval ZeroEval published resultsIBM Granite 3.0 Language Models technical report (GitHub)
448Llama 2 7B Chat · 8-bit Meta23±5*local~11GB4kOllama › Tulu 2 — base and instruction-tuned evaluations
449Qwen2-VL 2B Instruct · 4-bit Alibaba23±3*†local~3.9GB33kLM Studio › Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)
450StarCoderBase 15.5B · 8-bit BigCode23±2*†local~19GBllama.cpp › StarCoder 2 and The Stack v2: The Next Generation — authors' own runs
451Qwen2 1.5B Instruct · 8-bit Alibaba23±1local~4GB32kOllama › WildEval ZeroEval published resultsQwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card)Qwen2-0.5B-Instruct model card
452Granite 3.0 1B-A400M Instruct · 8-bit IBM23±3*†local~3GB4kllama.cpp › IBM Granite 3.0 Language Models technical report (GitHub)
453Code Llama 7B (base) · 8-bit Meta22±2local~9.1GB16kOllama › StarCoder 2 and The Stack v2: The Next Generation — authors' own runsMistral 7B — authors' own runsGranite Code Models: A Family of Open Foundation Models for Code Intelligence — authors' own runsTulu 2 — base and instruction-tuned evaluations
455Granite 4.0 H 350M · 4-bit IBM20±2†local~1.9GB33kLM Studio › ibm-granite/granite-4.0-h-1b model card
456Granite 4.0 350M · 16-bit IBM19±2local~2.9GB32kOllama › ibm-granite/granite-4.0-h-1b model cardEpoch AI — internal evaluations
457Llama 2 13B Chat · 8-bit Meta19±3*local~18GB4kOllama › Orca 2: Teaching Small Language Models How to Reason (Mitra et al., Microsoft)Tulu 2 — base and instruction-tuned evaluations
458Llama 2 7B (base) · 8-bit Meta18±2local~9.1GB4kllama.cpp › Mixtral of Experts — base-model evaluationsMistral 7B — authors' own runsMiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies (Hu et al.)Tulu 2 — base and instruction-tuned evaluations
460DeepSeek-Coder 1.3B Base · 8-bit DeepSeek17±3*local~3GB16kOllama › StarCoder 2 and The Stack v2: The Next Generation — authors' own runs
461StarCoderBase 7B · 4-bit BigCode17±2local~6GBOllama › StarCoder 2 and The Stack v2: The Next Generation — authors' own runsGranite Code Models: A Family of Open Foundation Models for Code Intelligence — authors' own runs
462StarCoderBase 3B · 4-bit BigCode13±3*†local~14GBhow to run › StarCoder 2 and The Stack v2: The Next Generation — authors' own runs
—DeepSeek LLM 7B Chat · q4_0 DeepSeeknot yet calibratedlocal~5.7GBOllama › DeepSeek vendor-reported benchmark scores for its own models (Hugging Face model cards)
—DeepSeek-V2 Lite Chat (16B-A2.4B) · Q8_0 DeepSeeknot yet calibratedlocal~19GB32kllama.cpp › DeepSeek vendor-reported benchmark scores for its own models (Hugging Face model cards)
—Eurus-2-7B-PRIME · Q8_0 PRIME-RLnot yet calibratedlocal~10GB4kllama.cpp › PRIME-RL Eurus-2-7B-PRIME model card, Evaluation table (own models)
—Eurus-2-7B-SFT · Q8_0 PRIME-RLnot yet calibratedlocal~10GBllama.cpp › PRIME-RL Eurus-2-7B-PRIME model card, Evaluation table (own models)
—Flan-T5 Base · fp32 Googlenot yet calibratedlocal~2.5GBhow to run › Scaling Instruction-Finetuned Language Models — authors' own runs
—Flan-T5 Large · fp32 Googlenot yet calibratedlocal~4.8GBhow to run › Scaling Instruction-Finetuned Language Models — authors' own runs
—Flan-T5 Small · fp32 Googlenot yet calibratedlocal~1.8GBhow to run › Scaling Instruction-Finetuned Language Models — authors' own runs
—Flan-T5 XL · fp32 Googlenot yet calibratedlocal~14GBhow to run › Scaling Instruction-Finetuned Language Models — authors' own runs
—Gemma 3n E2B IT · q8_0 Googlenot yet calibratedlocal~8.9GB32kOllama › Google DeepMind Gemma 3n E2B IT model card
—Gemma 3n E4B · q8_0 Googlenot yet calibratedlocal~12GB32kOllama › Google DeepMind Gemma 3n E2B IT model card
—GPT-2 XL (1.5B) · fp32 OpenAInot yet calibratedlocal~8.3GB1khow to run › EleutherAI GPT-Neo 2.7B model card — EleutherAI evaluation-harness runs
—GPT-Neo 1.3B · fp32 EleutherAInot yet calibratedlocal~7.1GB2khow to run › EleutherAI GPT-Neo 2.7B model card — EleutherAI evaluation-harness runs
—GPT-Neo 2.7B · fp32 EleutherAInot yet calibratedlocal~13GB2khow to run › EleutherAI GPT-Neo 2.7B model card — EleutherAI evaluation-harness runs
—Granite 4.1 30B · q4_k_m IBMnot yet calibratedlocal~19GBOllama › ibm-granite/granite-4.1-8b model card
—Granite 4.1 3B · q4_k_m IBMnot yet calibratedlocal~3.7GBOllama › ibm-granite/granite-4.1-8b model card
—Granite 4.1 8B · mlx-8bit IBMnot yet calibratedlocal~11GB131kLM Studio › ibm-granite/granite-4.1-8b model card
—K2 Horizon 7B · gguf-q8_0 Institute of Foundation Models (MBZUAI)not scored yetlocal~13GB524kllama.cpp ›
—Ministral 3 14B Reasoning 2512 · mlx-4bit Mistralnot yet calibratedlocal~10GB262kLM Studio › mistralai/Ministral-3-3B-Reasoning-2512 model cardmistralai/Ministral-3-8B-Reasoning-2512 model card
—Ministral 3 3B Reasoning 2512 · mlx-8bit Mistralnot yet calibratedlocal~5.3GB262kLM Studio › mistralai/Ministral-3-8B-Reasoning-2512 model cardmistralai/Ministral-3-3B-Reasoning-2512 model card
—Ministral 3 8B Reasoning 2512 · mlx-8bit Mistralnot yet calibratedlocal~11GB262kLM Studio › mistralai/Ministral-3-3B-Reasoning-2512 model cardmistralai/Ministral-3-8B-Reasoning-2512 model card
—Muse Glimmer 30B · q4_k_m Metanot scored yetlocal~21GB131kOllama ›
—OLMo 7B Instruct · bf16 Ai2not yet calibratedlocal~16GBhow to run › OLMo: Accelerating the Science of Language Models — authors' own runs
—OpenChat 3.5 · q8_0 OpenChatnot yet calibratedlocal~11GB8kOllama › OpenChat 3.5 model card (openchat/openchat_3.5)
—OpenHermes 2.5 Mistral 7B · mlx-bf16 Nous Researchnot yet calibratedlocal~17GBLM Studio › OpenChat 3.5 model card (openchat/openchat_3.5)
—Qwen2.5 0.5B Instruct · q8_0 Alibabanot yet calibratedlocal~2.6GB33kOllama › HuggingFaceTB/SmolLM2-360M-Instruct model card
—SmolLM2 360M Instruct · fp16 Hugging Facenot yet calibratedlocal~2.3GB8kOllama › HuggingFaceTB/SmolLM2-360M-Instruct model card

+ 296 more don't fit your machine or budget · show them

+ 1 models with fewer than 3 results left off, since one or two tests say little about overall ability · show them

509 of 511 catalogue models listed here · sources and licences

* new or thinly tested: ranked on too few benchmark results · † only the maker’s own numbers so far · ~RAM is estimated; full context may need more memory