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

Best AI models for agents to run locally on a 64GB Mac

A leaderboard of the best AI models for agents you can run locally on a Mac with 64GB of memory. Qwen3.8 27B is #1 right now, followed by Ornith 1.5 35B-A3B and Muse Glimmer 30B. Models are ranked on agent tests like SWE-bench and Terminal-Bench, where they use tools, write code and work in a terminal. Each model shows the download that fits and the command to run it.

153 models · scores combine 76 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 agent tasks evaluations only. Compare scores within this task. Models without task measurements are excluded.

Models for 64GB RAM · Apple Silicon → ~48GB fast · local only · reset
+ more
chip[M1 Max][M1 Ultra][M2 Max][M2 Ultra][M3 Max (16-core CPU / 40-core GPU)][M4 Pro][M4 Max (16-core CPU / 40-core GPU)][M5 Pro][M5 Max (40-core GPU)]
set[all sets][BFCL v3 · 374][BFCL v3 · 4][SWE-bench Multilingual · 4][SWE-bench Verified · 2][SWE-bench Multilingual · 2][BFCL V4 (December 2025 archive) · 2][τ²-bench Telecom · 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=64&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
46Qwen3.8 27B · 8-bit Alibaba71±1local~36GB262kOllama › LMArena leaderboard datasetQwen/Qwen3.8-27B 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)One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows (arXiv:2608.19741v4)
69Ornith 1.5 35B-A3B · 8-bit DeepReinforce65±2†local~41GB262kllama.cpp › ornith-ai/Ornith-1.5-35B-A3B model card
72Muse Glimmer 30B · 8-bit Meta65±2local~40GB131kOllama › meta-models/Muse-Glimmer-30B model card (speed and benchmark tables)Qwen/Qwen3.8-27B model card
76Qwen3.6 27B · 8-bit Alibaba64±1local~37GB262kOllama › Qwen/Qwen3.6-27B model cardQwen/Qwen3.8-27B model cardmeta-models/Muse-Glimmer-30B model card (speed and benchmark tables)One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows (arXiv:2608.19741v4)EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)
88KAT-Coder-V2.5-Dev (35B-A3B) · 8-bit Kuaishou (Kwaipilot)62±2†local~40GB262kllama.cpp › Kwaipilot/KAT-Coder-V2.5-Dev model card
89Qwen3.5 27B · 8-bit Alibaba62±1local~37GB262kOllama › Qwen/Qwen3.5-122B-A10B model cardKwaipilot/KAT-Coder-V2.5-Dev model cardQwen/Qwen3.6-35B-A3B model cardQwen/Qwen3.6-27B model cardQwen/Qwen3.5-27B model cardContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)
93Qwen3.6 35B-A3B · 4-bit Alibaba61±1local~41GB262kLM Studio › Kwaipilot/KAT-Coder-V2.5-Dev model cardQwen/Qwen3.6-35B-A3B model cardQwen/Qwen3.6-27B model cardK-BrowseComp: A Web Browsing Agent Benchmark Grounded in Korean Contexts (arXiv:2606.02404v1)
98Ornith 1.0 35B (35B-A3B) · 4-bit DeepReinforce60±1local~41GB262kLM Studio › ornith-ai/Ornith-1.5-35B-A3B model cardKwaipilot/KAT-Coder-V2.5-Dev model card
100Qwen3.5 35B-A3B · 8-bit Alibaba60±1local~47GB262kOllama › Qwen/Qwen3.5-122B-A10B model cardQwen/Qwen3.6-35B-A3B model cardKwaipilot/KAT-Coder-V2.5-Dev model cardQwen/Qwen3.5-27B model cardEEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)
103SmolLM3 3B · 8-bit Hugging Face59±11*†local~5.9GB66kllama.cpp › HuggingFaceTB/SmolLM3-3B model card (Evaluation section)
109Mistral Medium 3.5 (128B) · 2-bit Mistral58±2local~42GB262kllama.cpp › LMArena leaderboard datasetSolar Open 2 Technical Report (Upstage, arXiv:2607.20062v2)EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)
114Gemma 4 31B · 4-bit Google57±1local~37GB262kLM Studio › Gemma Team — Gemma 4 Technical ReportKwaipilot/KAT-Coder-V2.5-Dev model cardmeta-models/Muse-Glimmer-30B model card (speed and benchmark tables)ContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)Qwen/Qwen3.6-27B model cardgoogle/gemma-4-31B-it model card+4
119Spark-X2.5 4B · 16-bit iFLYTEK (SparkLLM)56±2†local~10GB1MOllama › XHToken/Spark-X2.5-4B model card
123EXAONE 4.5 33B · 8-bit LG AI Research56±3*†local~45GB262kllama.cpp › LG AI Research vendor-reported benchmark scores for its own EXAONE models (Hugging Face model cards)
127MiMo-V2.6-Distill-Qwen-9B · 8-bit Xiaomi56±3†local~12GBllama.cpp › XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B model card
137MiniCPM5 2B · 4-bit OpenBMB55±3†local~4.3GB131kLM Studio › openbmb/MiniCPM5-2B model card
142Granite 4.2 30B · 8-bit IBM54±2†local~39GB131kOllama › ibm-granite/granite-4.2-30b model card
148Nemotron-Labs-3 Puzzle 75B-A9B · 2-bit NVIDIA53±2†local~35GBllama.cpp › NVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards)
149Seed-OSS 36B Instruct · 8-bit ByteDance Seed52±3†local~48GB524kllama.cpp › ByteDance-Seed/Seed-OSS-36B-Instruct model card
151Mistral Small 4 (119B-A6.5B) · 2-bit Mistral52±3*local~40GB262kllama.cpp › ContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)
155GLM-4.5-Air · 2-bit Z.ai51±2local~35GB131kllama.cpp › GLM-4.5 Team (Zhipu AI & Tsinghua) — GLM-4.5 technical reportSolar Open Technical Report (arXiv:2601.07022v1)Berkeley Function Calling Leaderboard result archiveGalileo Agent Leaderboard (v1 and v2 results)
156Jamba 1.5 Mini · 4-bit AI21 Labs51±10*†local~34GBllama.cpp › Jamba-1.5: Hybrid Transformer-Mamba Models at Scale (AI21, arXiv 2408.12570v1)
157Qwen3.5 9B · 4-bit Alibaba51±1local~12GB262kLM Studio › JetBrains/Mellum2.1-12B-A2.5B-Thinking model cardXiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B model cardXHToken/Spark-X2.5-4B model cardContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)Qwen/Qwen3.5-9B model cardJetBrains/Mellum2-12B-A2.5B-Instruct model card+3
159Mellum2.1 12B-A2.5B Thinking · 8-bit JetBrains50±3†local~15GBllama.cpp › JetBrains/Mellum2.1-12B-A2.5B-Thinking model card
162Qwen3-Next 80B-A3B Thinking · 2-bit Alibaba49±2†local~27GB131kllama.cpp › Qwen/Qwen3-Next-80B-A3B-Thinking model cardQwen/Qwen3.5-9B model card
163Command A Reasoning (08-2025) · 2-bit Cohere49±3local~34GBllama.cpp › Berkeley Function Calling Leaderboard result archiveCohere vendor-reported benchmark scores for its own models (cohere.com blog)
164Granite 4.1 30B · 4-bit IBM49±5*†local~19GBOllama › ibm-granite/granite-4.1-8b model card
166Gemma 4 12B · 8-bit Google48±2local~15GB262kllama.cpp › Gemma Team — Gemma 4 Technical ReportXHToken/Spark-X2.5-4B model cardgoogle/gemma-4-31B-it model card
168Gemma 4 26B-A4B · 4-bit Google48±1local~31GB262kLM Studio › Gemma Team — Gemma 4 Technical ReportKwaipilot/KAT-Coder-V2.5-Dev model cardContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)google/gemma-4-31B-it model cardgoogle/diffusiongemma-26B-A4B-it model cardQwen/Qwen3.6-35B-A3B model card+1
170Qwen3-VL 32B Thinking · 8-bit Alibaba48±1local~39GB256kOllama › 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)HyperCLOVA X 32B Think (NAVER Cloud, arXiv:2601.03286v1)
171Granite 4.2 8B · 8-bit IBM48±2†local~11GB131kllama.cpp › ibm-granite/granite-4.2-30b model card
174Qwen3.5 4B · 4-bit Alibaba47±1local~6.9GB262kLM Studio › openbmb/MiniCPM5-2B model cardXHToken/Spark-X2.5-4B model cardQwen/Qwen3.5-9B model cardJetBrains/Mellum2-12B-A2.5B-Instruct model cardLiquidAI/LFM2.5-2.6B model cardJetBrains/Mellum2-12B-A2.5B-Thinking model card+1
177Qwen3-VL 32B Instruct · 8-bit Alibaba46±3*†local~39GB256kOllama › 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)
180Solar Open 100B (102B-A12B) · 2-bit Upstage46±3†local~41GB131kllama.cpp › Solar Open 2 Technical Report (Upstage, arXiv:2607.20062v2)Solar Open Technical Report (arXiv:2601.07022v1)
182LFM2.5 2.6B · 4-bit Liquid AI46±2local~4.5GB131kLM Studio › openbmb/MiniCPM5-2B model cardLiquidAI/LFM2.5-2.6B model card
185Qwen3 4B Thinking 2507 · 8-bit Alibaba46±2†local~7.1GB262kOllama › Qwen/Qwen3-4B-Thinking-2507 model card
188Nemotron-Cascade-14B-Thinking · 8-bit NVIDIA46±4*†local~18GBllama.cpp › NVIDIA — Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models
189Qwen3 14B · 8-bit Alibaba46±1local~21GB33kOllama › Berkeley Function Calling Leaderboard result archiveMUA-RL: Multi-turn User-interacting Agent Reinforcement Learning for agentic tool use (arXiv:2508.18669v1)
191Qwen3 Coder 30B-A3B Instruct · 8-bit Alibaba45±2local~41GB160kOllama › Kwaipilot/KAT-Coder-V2.5-Dev model cardQwen/Qwen3-Coder-30B-A3B-Instruct model card
194Qwen3 30B-A3B Thinking 2507 · 8-bit Alibaba45±1local~41GB262kOllama › Qwen/Qwen3-Next-80B-A3B-Thinking model cardQwen/Qwen3-30B-A3B-Thinking-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)ByteDance-Seed/Seed-OSS-36B-Instruct model cardQwen/Qwen3.5-9B model cardLiquidAI/LFM2.5-8B-A1B model card
195Nemotron 3 Nano 30B-A3B · 8-bit NVIDIA45±3*†local~42GB262kOllama › NVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards)
196Granite 4.2 3B · 8-bit IBM45±2local~5.6GB131kllama.cpp › openbmb/MiniCPM5-2B model cardibm-granite/granite-4.2-30b model card
197Mellum2 12B-A2.5B Instruct · 8-bit JetBrains45±4*†local~15GB131kllama.cpp › JetBrains/Mellum2-12B-A2.5B-Instruct model card
198Nemotron Nano 9B v2 · 8-bit NVIDIA45±4*†local~13GB131kllama.cpp › NVIDIA — Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning ModelsNVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards)
200Qwen3-VL 4B Thinking · 8-bit Alibaba45±3*†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)
202EXAONE 4.0 32B · 8-bit LG AI Research45±1local~42GB131kllama.cpp › HyperCLOVA X 32B Think (NAVER Cloud, arXiv:2601.03286v1)LG AI Research vendor-reported benchmark scores for its own EXAONE models (Hugging Face model cards)
203Qwen3-VL 30B-A3B Thinking · 8-bit Alibaba44±2*†local~37GB256kOllama › 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)
204Qwen3-VL 30B-A3B Instruct · 8-bit Alibaba44±3*†local~37GB256kOllama › 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)
205Qwen3 32B · 8-bit Alibaba44±1local~43GB131kOllama › Berkeley Function Calling Leaderboard result archiveQwen/Qwen3-Next-80B-A3B-Thinking model cardQwen/Qwen3-Next-80B-A3B-Instruct 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)Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time ScalingMUA-RL: Multi-turn User-interacting Agent Reinforcement Learning for agentic tool use (arXiv:2508.18669v1)+1
206Command A (03-2025) · 2-bit Cohere44±1local~37GB256kllama.cpp › Berkeley Function Calling Leaderboard result archiveCohere — Command A technical reportFinance Agent Benchmark (Vals AI, arXiv:2508.00828v1)
207Falcon-H1R 7B · 8-bit TII44±2*†local~12GB262kllama.cpp › Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time Scaling
209Qwen3 8B · 8-bit Alibaba44±1local~12GB33kOllama › Berkeley Function Calling Leaderboard result archiveFalcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time ScalingSynthesize and Reward: Reinforcement Learning for Multi-Step Tool Use in Live Environments (arXiv:2606.03892v2)MUA-RL: Multi-turn User-interacting Agent Reinforcement Learning for agentic tool use (arXiv:2508.18669v1)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)EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)+1
212Qwen3-Next 80B-A3B Instruct · 2-bit Alibaba43±2†local~27GB131kllama.cpp › Qwen/Qwen3-Next-80B-A3B-Instruct model card
213Qwen3-VL 8B Instruct · 4-bit Alibaba43±3*local~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
215Nemotron-Cascade-8B · 8-bit NVIDIA43±4*†local~11GBllama.cpp › NVIDIA — Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models
217Apriel 1.6 15B Thinker · 8-bit ServiceNow43±4†local~21GBllama.cpp › ServiceNow-AI/Apriel-1.6-15b-Thinker model card
218Spark-X2.5 1.7B · 16-bit iFLYTEK (SparkLLM)43±2†local~5.1GB1MOllama › XHToken/Spark-X2.5-4B model card
219Qwen2.5 72B Instruct · 4-bit Alibaba43±1local~48GB33kllama.cpp › Berkeley Function Calling Leaderboard result archiveGalileo Agent Leaderboard (v1 and v2 results)ContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)Cohere — Command A technical reportMulti-SWE-bench experiments (ByteDance Seed baseline runs)
220Qwen2.5 32B Instruct · 8-bit Alibaba43±1*local~43GB131kOllama › Berkeley Function Calling Leaderboard result archive
222Qwen3-VL 4B Instruct · 8-bit Alibaba42±3*†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)
223Qwen3 30B-A3B Instruct 2507 · 8-bit Alibaba42±1local~41GB262kOllama › Qwen/Qwen3-30B-A3B-Instruct-2507 model cardQwen/Qwen3-Next-80B-A3B-Instruct model cardBerkeley Function Calling Leaderboard result archive
226Qwen3-VL 8B Thinking · 4-bit Alibaba42±3*†local~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)
227gpt-oss-20b · 4-bit OpenAI42±2local~25GB131kLM Studio › OpenAI — gpt-oss-120b & gpt-oss-20b Model CardServiceNow BrowserGym leaderboard (AgentLab GenericAgent runs)Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces — benchmark authors' own agent-model trialsMultiverse Computing HyperNova 60B 2602 model card: evaluation table (own runs incl. gpt-oss baselines)T1-Bench: Benchmarking Multi-Scenario Agents in Real-World Domains (arXiv:2606.11070v1)
230Ministral 3 14B · 8-bit Mistral42±3local~20GB262kOllama › JetBrains/Mellum2-12B-A2.5B-Instruct model cardContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)
231Qwen3 30B-A3B (Apr 2025, hybrid) · 8-bit Alibaba42±1local~41GB33kOllama › Berkeley Function Calling Leaderboard result archiveQwen/Qwen3-4B-Instruct-2507 model cardQwen/Qwen3-30B-A3B-Instruct-2507 model cardQwen/Qwen3-30B-A3B-Thinking-2507 model cardQwen/Qwen3-4B-Thinking-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)
233Granite 4.1 8B · 4-bit IBM41±2*†local~11GB131kLM Studio › Synthesize and Reward: Reinforcement Learning for Multi-Step Tool Use in Live Environments (arXiv:2606.03892v2)ibm-granite/granite-4.1-8b model card
235Granite 4.1 3B · 4-bit IBM41±5*†local~3.7GBOllama › ibm-granite/granite-4.1-8b model card
239Ministral 3 8B · 8-bit Mistral41±4*local~13GB262kOllama › ContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)
240Mistral Small 3.2 24B · 8-bit Mistral41±3*local~32GB128kOllama › Berkeley Function Calling Leaderboard result archiveGalileo Agent Leaderboard (v1 and v2 results)
242LFM2.5 8B-A1B · 8-bit Liquid AI41±2local~13GB128kOllama › openbmb/MiniCPM5-2B model cardLiquidAI/LFM2.5-8B-A1B model card
243Qwen3 4B · 8-bit Alibaba41±1local~7.1GB33kOllama › Berkeley Function Calling Leaderboard result archiveQwen/Qwen3-4B-Instruct-2507 model cardQwen/Qwen3-4B-Thinking-2507 model cardEEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)HuggingFaceTB/SmolLM3-3B model card (Evaluation section)
244Qwen2.5 14B Instruct · 8-bit Alibaba40±1*local~21GB131kOllama › Berkeley Function Calling Leaderboard result archiveReinforcing Real-world Service Agents: InteractCS-RL (arXiv:2602.22697v1)
245LFM2.5 1.2B Thinking · 8-bit Liquid AI40±4*†local~3.5GB33kOllama › LiquidAI/LFM2.5-1.2B-Thinking model card
249Mistral Small 3.1 24B (2503) · 8-bit Mistral40±1local~32GB128kOllama › Berkeley Function Calling Leaderboard result archiveGalileo Agent Leaderboard (v1 and v2 results)Finance Agent Benchmark (Vals AI, arXiv:2508.00828v1)
251Mistral Large 2 (2407) · 2-bit Mistral40±1*local~45GB128kllama.cpp › Berkeley Function Calling Leaderboard result archiveJamba-1.5: Hybrid Transformer-Mamba Models at Scale (AI21, arXiv 2408.12570v1)
252QwQ-32B · 8-bit Alibaba40±1*local~43GB131kOllama › Berkeley Function Calling Leaderboard result archive
253Qwen3 4B Instruct 2507 · 8-bit Alibaba40±1local~7.1GB262kOllama › Qwen/Qwen3-4B-Instruct-2507 model cardSynthesize and Reward: Reinforcement Learning for Multi-Step Tool Use in Live Environments (arXiv:2606.03892v2)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)Berkeley Function Calling Leaderboard result archive
259Gemma 4 E4B · 4-bit Google39±1local~11GB131kLM Studio › Gemma Team — Gemma 4 Technical Reportopenbmb/MiniCPM5-2B model cardJetBrains/Mellum2.1-12B-A2.5B-Thinking model cardXHToken/Spark-X2.5-4B model cardLiquidAI/LFM2.5-2.6B model cardContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)+3
260Phi-3.5-mini-instruct · 8-bit Microsoft39±11*†local~6.8GB128kOllama › Microsoft — Phi-4-Mini Technical Report
261Qwen3-VL 2B Thinking · 8-bit Alibaba39±3*†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)
264Qwen2.5 7B Instruct · 8-bit Alibaba38±1local~12GB131kOllama › Berkeley Function Calling Leaderboard result archiveSynthesize and Reward: Reinforcement Learning for Multi-Step Tool Use in Live Environments (arXiv:2606.03892v2)Reinforcing Real-world Service Agents: InteractCS-RL (arXiv:2602.22697v1)allenai/Olmo-3-7B-Instruct model card
265Granite 4.0 1B · 16-bit IBM38±3*local~6.1GB128kOllama › LiquidAI/LFM2.5-1.2B-Thinking model cardibm-granite/granite-4.0-h-1b model card
268Qwen3 1.7B · 8-bit Alibaba38±1local~4.6GB33kOllama › Berkeley Function Calling Leaderboard result archiveQwen (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/LFM2.5-1.2B-Thinking model cardEEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)HuggingFaceTB/SmolLM3-3B model card (Evaluation section)
269Qwen3-VL 2B Instruct · 8-bit Alibaba38±3*†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)
271Gemma 3 27B · 8-bit Google38±1local~37GB131kOllama › Berkeley Function Calling Leaderboard result archiveGemma Team — Gemma 4 Technical Reportgoogle/gemma-4-31B-it model card
272Ministral 3 14B Reasoning 2512 · 4-bit Mistral38±5*local~10GB262kLM Studio › JetBrains/Mellum2-12B-A2.5B-Thinking model card
277Llama 3.1 70B Instruct · 4-bit Meta37±1local~46GB128kOllama › Berkeley Function Calling Leaderboard result archiveServiceNow BrowserGym leaderboard (AgentLab GenericAgent runs)MASSIVE-Agents — Kulkarni et al. (2025)
278Nemotron 3 Nano 4B · 8-bit NVIDIA37±2local~7GB262kOllama › openbmb/MiniCPM5-2B model cardNVIDIA vendor-reported benchmark scores for its own Nemotron models (Hugging Face model cards)
280Sky-T1 32B Preview · 4-bit NovaSky (UC Berkeley)37±1*local~38GBLM Studio › Berkeley Function Calling Leaderboard result archive
281Granite 4.0 H 1B · 8-bit IBM37±3*local~3.2GB131kOllama › LiquidAI/LFM2.5-1.2B-Thinking model cardibm-granite/granite-4.0-h-1b model card
282Llama 3.3 70B Instruct · 4-bit Meta36±1local~46GB128kOllama › Berkeley Function Calling Leaderboard result archiveGalileo Agent Leaderboard (v1 and v2 results)Cohere — Command A technical reportContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)Finance Agent Benchmark (Vals AI, arXiv:2508.00828v1)EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)
284LFM2.5 1.2B Instruct · 4-bit Liquid AI36±4*†local~2.8GB33kLM Studio › LiquidAI/LFM2.5-1.2B-Thinking model card
285Command R7B (12-2024) · 8-bit Cohere36±1local~12GB128kOllama › Berkeley Function Calling Leaderboard result archiveCohere — Command A technical report
286Mellum2 12B-A2.5B Thinking · 8-bit JetBrains36±3†local~15GB131kllama.cpp › JetBrains/Mellum2.1-12B-A2.5B-Thinking model cardJetBrains/Mellum2-12B-A2.5B-Thinking model card
287Gemma 3 12B · 8-bit Google36±1*local~17GB131kOllama › Berkeley Function Calling Leaderboard result archive
288Llama 3 70B Instruct · 4-bit Meta36±1*local~47GB8kOllama › Berkeley Function Calling Leaderboard result archive
290Ministral 8B (2410) · 8-bit Mistral36±1*local~10GB128kllama.cpp › Berkeley Function Calling Leaderboard result archiveGalileo Agent Leaderboard (v1 and v2 results)
291Qwen2.5 3B Instruct · 8-bit Alibaba36±1*local~5.9GB33kOllama › Berkeley Function Calling Leaderboard result archive
292Mistral Small Instruct 22B (2409) · 8-bit Mistral AI36±4*local~30GB32kOllama › Galileo Agent Leaderboard (v1 and v2 results)
293Mistral NeMo Instruct (2407) · 8-bit Mistral35±1local~18GB128kOllama › Berkeley Function Calling Leaderboard result archiveGalileo Agent Leaderboard (v1 and v2 results)
294Granite 3.1 8B Instruct · 8-bit IBM35±4*local~12GB128kOllama › Berkeley Function Calling Leaderboard result archive
295MiniCPM3-4B · 4-bit OpenBMB35±1*local~4.1GB32kllama.cpp › Berkeley Function Calling Leaderboard result archive
296Granite 3.2 8B Instruct · 8-bit IBM35±4*local~12GB131kOllama › Berkeley Function Calling Leaderboard result archive
297Llama 3.1 8B Instruct · 8-bit Meta35±1local~12GB128kOllama › Berkeley Function Calling Leaderboard result archiveGalileo Agent Leaderboard (v1 and v2 results)MASSIVE-Agents — Kulkarni et al. (2025)
298Command R Plus (04-2024) · 2-bit Cohere34±1*local~43GB128kllama.cpp › Berkeley Function Calling Leaderboard result archive
299DeepSeek R1 0528 Qwen3 8B · 8-bit DeepSeek34±5*local~12GB128kOllama › Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time Scaling
300Qwen2.5 1.5B Instruct · 8-bit Alibaba34±1*local~4GB33kOllama › Berkeley Function Calling Leaderboard result archive
301Phi-4 14B · 8-bit Microsoft34±1*local~21GB16kOllama › Berkeley Function Calling Leaderboard result archive
303Qwen3 0.6B · 8-bit Alibaba34±1*local~3GB33kOllama › Berkeley Function Calling Leaderboard result archive
304Hermes 2 Pro Llama 3 8B · 4-bit Nous Research34±1*local~18GB8khow to run › Berkeley Function Calling Leaderboard result archive
305Granite 4.0 H 350M · 4-bit IBM34±5*†local~1.9GB33kLM Studio › ibm-granite/granite-4.0-h-1b model card
307EXAONE 4.0 1.2B · 8-bit LG AI Research33±3†local~3.6GB66kllama.cpp › LG AI Research vendor-reported benchmark scores for its own EXAONE models (Hugging Face model cards)
308Falcon3 10B Instruct · 8-bit TII33±1*local~15GB32kOllama › Berkeley Function Calling Leaderboard result archive
309Llama 4 Scout (109B-A17B) · 2-bit Meta33±1local~36GB328kllama.cpp › Berkeley Function Calling Leaderboard result archiveGalileo Agent Leaderboard (v1 and v2 results)Finance Agent Benchmark (Vals AI, arXiv:2508.00828v1)
310Llama 3.2 3B Instruct · 8-bit Meta33±1local~6.1GB131kOllama › Berkeley Function Calling Leaderboard result archiveMASSIVE-Agents — Kulkarni et al. (2025)
312Gemma 4 E2B · 4-bit Google32±2local~7.7GB131kLM Studio › Gemma Team — Gemma 4 Technical Reportopenbmb/MiniCPM5-2B model cardXHToken/Spark-X2.5-4B model cardLiquidAI/LFM2.5-2.6B model cardgoogle/gemma-4-31B-it model cardContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)+2
313Falcon3 7B Instruct · 8-bit TII32±1*local~11GB32kOllama › Berkeley Function Calling Leaderboard result archive
314Qwen3.5 2B · 8-bit Alibaba32±2local~4.9GB262kOllama › XHToken/Spark-X2.5-4B model cardopenbmb/MiniCPM5-2B model cardQwen/Qwen3.5-2B model card
315Magistral Small 1.0 (2506) · 8-bit Mistral31±7*local~32GB128kOllama › Galileo Agent Leaderboard (v1 and v2 results)
316Mixtral 8x7B Instruct v0.1 · 4-bit Mistral31±1*local~31GB32kOllama › Berkeley Function Calling Leaderboard result archive
318Qwen2 7B Instruct · 8-bit Alibaba31±1*local~12GB131kOllama › Berkeley Function Calling Leaderboard result archive
319Bielik 11B v2.3 Instruct · 4-bit SpeakLeash30±1*local~14GBLM Studio › Berkeley Function Calling Leaderboard result archive
320Command R (08-2024) · 8-bit Cohere30±8*†local~43GB128kOllama › Cohere — Command A technical report
321Ministral 3 3B · 8-bit Mistral30±6*local~6.9GB262kOllama › ContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)
322ChatGLM3-6B · 8-bit Zhipu AI (THUDM)30±10*†local~8.4GBllama.cpp › ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools (Team GLM, arXiv 2406.12793v2)
323GLM-4 9B Chat · 8-bit Zhipu AI (Z.ai)30±1local~14GB128kOllama › Berkeley Function Calling Leaderboard result archiveChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools (Team GLM, arXiv 2406.12793v2)
324Gemma 3 4B · 8-bit Google30±1*local~7.4GB131kOllama › Berkeley Function Calling Leaderboard result archive
326Granite 4.0 350M · 16-bit IBM30±3*local~2.9GB32kOllama › Berkeley Function Calling Leaderboard result archiveibm-granite/granite-4.0-h-1b model card
327Olmo 3 7B Instruct · 8-bit Ai229±4local~11GB66kOllama › JetBrains/Mellum2-12B-A2.5B-Instruct model cardallenai/Olmo-3-7B-Instruct model card
329NexusRaven V2 13B · 8-bit Nexusflow28±1*local~16GB16kllama.cpp › Berkeley Function Calling Leaderboard result archive
330Qwen2.5 0.5B Instruct · 8-bit Alibaba28±1*local~2.6GB33kOllama › Berkeley Function Calling Leaderboard result archive
331Qwen3.5 0.8B · 8-bit Alibaba27±5*†local~3.1GB262kOllama › Qwen/Qwen3.5-2B model card
332Llama 3 8B Instruct · 8-bit Meta26±1*local~12GB8kOllama › Berkeley Function Calling Leaderboard result archiveChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools (Team GLM, arXiv 2406.12793v2)
333Falcon3 3B Instruct · 8-bit TII26±1*local~6.1GB32kOllama › Berkeley Function Calling Leaderboard result archive
334Phi-4-mini 3.8B · 8-bit Microsoft23±1local~6.9GB128kOllama › Berkeley Function Calling Leaderboard result archiveMicrosoft — Phi-4-Mini Technical Report
335Qwen2 1.5B Instruct · 8-bit Alibaba23±2*local~4GB32kOllama › Berkeley Function Calling Leaderboard result archive
336Llama 3.2 1B Instruct · 4-bit Meta18±1local~5.2GB131kOllama › Berkeley Function Calling Leaderboard result archiveLiquidAI/LFM2.5-1.2B-Thinking model cardMASSIVE-Agents — Kulkarni et al. (2025)
337DeepSeek-Coder-V2 Lite Instruct (16B-A2.4B) · 8-bit DeepSeek18±2*local~22GB128kOllama › Berkeley Function Calling Leaderboard result archive
338QwQ-32B-Preview · 8-bit Alibaba17±2*local~43GB33kOllama › Berkeley Function Calling Leaderboard result archive
339Falcon3 1B Instruct · 8-bit TII17±2*local~4.1GB8kOllama › Berkeley Function Calling Leaderboard result archive
340Gemma 3 1B · 4-bit Google13±2local~3GB32kLM Studio › Berkeley Function Calling Leaderboard result archiveLiquidAI/LFM2.5-1.2B-Thinking model card
—Devstral Small 2 24B · q8_0 Mistralnot yet calibratedlocal~32GB262kOllama › mistralai/Devstral-Small-2-24B-Instruct-2512 model card
—GLM-4.7-Flash · Q8_0 Z.ainot scored yetlocal~35GB200kllama.cpp ›
—Granite 4.0 H Micro 3B · mlx-8bit IBMnot yet calibratedlocal~5.1GB131kLM Studio › ibm-granite/granite-4.0-h-small model card
—Granite 4.0 H Small (32B-A9B) · Q8_0 IBMnot yet calibratedlocal~38GB131kllama.cpp › ibm-granite/granite-4.0-h-small model card
—Granite 4.0 H Tiny (7B-A1B) · mlx-8bit IBMnot yet calibratedlocal~9.3GB131kLM Studio › ibm-granite/granite-4.0-h-small model card
—Granite 4.0 Micro 3B · Q8_0 IBMnot yet calibratedlocal~5.3GB131kllama.cpp › ibm-granite/granite-4.0-h-small model card
—K2 Horizon 7B · gguf-q8_0 Institute of Foundation Models (MBZUAI)not scored yetlocal~13GB524kllama.cpp ›
—Laguna XS 2.1 (33B-A3B) · q8_0 Poolsidenot yet calibratedlocal~39GB262kOllama › Poolside vendor-reported benchmark scores for its own Laguna models (Hugging Face model cards)
—Laguna XS.2 · mlx-8bit Poolsidenot yet calibratedlocal~39GB262kLM Studio › Poolside vendor-reported benchmark scores for its own Laguna models (Hugging Face model cards)Poolside — Laguna M.1/XS.2 Technical Report
—Llama 3.3 Nemotron Super 49B v1 · q4_k_m NVIDIAnot scored yetlocal~33GBllama.cpp ›
—Nemotron Cascade 2 30B-A3B · Q8_0 NVIDIAnot scored yetlocal~37GB256kllama.cpp ›
—Qwen3-Coder-Next (80B-A3B) · ud-q2_k_xl Alibabanot scored yetlocal~27GB262kllama.cpp ›

+ 209 more don't fit your machine or budget · show them

+ 30 models with fewer than 3 results left off, since one or two tests say little about overall ability · show them

362 of 398 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