The local LLM leaderboard
According to the LLMCheck index (18 November 2026), MiMo-V2.6-Flash and Qwen 3.6-35B-A3B tie for #1 of 130 open-weight models at Score 81 (MiMo-V2.6-Flash is listed first on the alphabetical tie-break). MiMo-V2.6-Flash needs a Mac Studio with 192 GB or more; Qwen 3.6-35B-A3B needs 24 GB, the best pick at 32 GB or less. Every speed carries a provenance mark.
| # | Model | Score↓0–100 composite | Speedtok/s decode | RAM @64Kagent context | RAMunified min | License |
|---|---|---|---|---|---|---|
| 1 | MiMo-V2.6-Flash 309B MoE · MXFP4 · Xiaomi |
81 |
community77 M5 Ultracommunityprefill 2,755 @8KcommunityTTFT 0.609 s @1K | estimated172 GB | 170 GB | MIT |
| 2 | Qwen 3.6-35B-A3B 35B MoE · 4-bit · Alibaba ollama run qwen3.6:35b-a3bHugging Face |
81 |
community130communityprefill 4,576 @1KcommunityTTFT 0.224 s @1K | estimated26 GB | 24 GB | Apache 2.0 |
| 3 | GLM-4.7-Flash 31B MoE · Q4_K_M · Zhipu AI ollama run glm-4.7-flashHugging Face |
80 |
estimated107 | estimated22 GB | 18 GB | MIT |
| 4 | Recursive Flint 39.8B MoE · 4-bit · Recursive ollama run flint:40bHugging Face |
80 |
community123communityprefill 3,592 @8KcommunityTTFT 0.309 s @1K | estimated22 GB | 20 GB | Apache 2.0 |
| 5 | KAT-Coder-V2.5 35B MoE · OptiQ-4bit · Kwaipilot |
79 |
community89communityprefill 4,018 @4KcommunityTTFT 0.316 s @1K | estimated26 GB | 24 GB | Apache 2.0 |
| 6 | Nemotron 3.5 Lightning 30B MoE · Q4_0 · NVIDIA ollama run nemotron-3.5-lightning:30bHugging Face |
79 |
community117 | estimated19 GB | 18 GB | OpenMDW |
| 7 | North Mini Code 1.0 30.5B MoE · 4-bit · Cohere ollama run north-mini-code-1.0Hugging Face |
79 |
community120communityprefill 3,424 @8KcommunityTTFT 0.376 s @1K | estimated23 GB | 21 GB | Apache 2.0 |
| 8 | Gemma 4 26B-A4B 26B MoE · Q4_K_M · Google ollama run gemma4:26b-a4bHugging Face |
78 |
sourced107communityprefill 3,852 @4K | estimated20 GB | 18 GB | Apache 2.0 |
| 9 | MiniCPM-V 4.7 35B-A3B 35.2B MoE · MXFP4 · OpenBMB ollama run openbmb/minicpm-v4.7:35b-a3bHugging Face |
78 |
community100communityprefill 3,583 @4KcommunityTTFT 0.804 s @1K | estimated25 GB | 23 GB | Apache 2.0 |
| 10 | Raon-1 mini preview 28.8B MoE · Q4_K_M · KRAFTON |
77 |
estimated123 | estimated20 GB | 18 GB | Apache 2.0 |
| 11 | LFM2.5-24B-A2B 24B MoE · Q4_K_M · Liquid AI ollama run lfm2.5:24bHugging Face |
76 |
sourced184communityprefill 3,772 @4K | estimated19 GB | 17 GB | LFM Open |
| 12 | Laguna XS 2.1 33B MoE · Q4_K_M · Poolside |
76 |
estimated117 | estimated27 GB | 24 GB | OpenMDW |
| 13 | Bonsai 27B 27B · Q1_0_g128 · Prism ML |
72 |
sourced66 | estimated11 GB | 6 GB | Apache 2.0 |
| 14 | Qwen3.8-27B 27.8B · 4-bit · Alibaba ollama run qwen3.8:27bHugging Face |
72 |
community31communityprefill 839 @4K | estimated24 GB | 19 GB | Apache 2.0 |
| 15 | Nemotron-Cascade 2 30B MoE · Q4_K_M · NVIDIA ollama run nemotron-cascade-2:30bHugging Face |
71 |
estimated117 | estimated19 GB | 18 GB | NVIDIA Open |
| 16 | DeepSeek V4 Flash 284B MoE · Q2_K · DeepSeek |
70 |
community39communityTTFT 0.4 s | estimated121 GB | 120 GB | MIT |
| 17 | DeepSeek V4 Flash Vision-Exp 305B MoE · Q2_K · DeepSeek |
70 |
estimated39 | estimated121 GB | 120 GB | MIT |
| 18 | DeepSeek V4.1 Flash 552B MoE · oQ4e · DeepSeek |
70 |
estimated29 M5 Ultra | estimated435 GB | 434 GB | MIT |
| 19 | K2-Horizon-3.7B 5.1B · Q4_K_M · MBZUAI IFM |
70 |
estimated133 | estimated14 GB | 4 GB | Apache 2.0 |
| 20 | Ling-3.0-flash-VL 124B MoE · Q4_K_M · Ant Group |
70 |
estimated53 | estimated57 GB | 56 GB | MIT |
| 21 | Maple Preview 20B-A1B 20B MoE · Ternary · DeepGrove |
70 |
sourced387 M5 Pro | estimated7 GB | 6 GB | MIT |
| 22 | Nanbeige4.3-33B 33.1B · oQ4e · Nanbeige |
70 |
community25communityprefill 613 @4KcommunityTTFT 1.665 s @1K | estimated31 GB | 17 GB | Apache 2.0 |
| 23 | Nemotron 3.5 Bolt 74B-A9B 74B MoE · 5-bit · NVIDIA |
70 |
community64communityprefill 1,332 @4K | estimated46 GB | 45 GB | OpenMDW |
| 24 | Qwen 3.5 35B-A3B 35B MoE · Q4_K_M · Alibaba ollama run qwen3.5:35b-a3bHugging Face |
70 |
estimated136 | estimated26 GB | 24 GB | Apache 2.0 |
| 25 | Qwen 3.6-27B 27B · Q4_K_M · Alibaba ollama run qwen3.6:27bHugging Face |
70 |
estimated30 | estimated23 GB | 18 GB | Apache 2.0 |
| 26 | Qwen3-Coder-Next 80B MoE · Q4_K_M · Alibaba ollama run qwen3-coder-nextHugging Face |
70 |
estimated84 | estimated50 GB | 48 GB | Apache 2.0 |
| 27 | GLM 5.2 753B MoE · Q4_K_M · Zhipu AI |
69 |
estimated22 M5 Ultra | estimated262 GB | 256 GB | MIT |
| 28 | MiniCPM5-2B 2.52B · Q4_K_M · OpenBMB |
69 |
estimated225 | estimated6 GB | 3 GB | Apache 2.0 |
| 29 | Ternary Bonsai 2 27B 27.36B · PQ2_0 · Prism ML ollama run bonsai-2:27bHugging Face |
69 |
sourced47 | estimated14 GB | 9 GB | Apache 2.0 |
| 30 | GLM-5.3-Flash 320B MoE · IQ3_XXS · Zhipu AI |
68 |
estimated27 | estimated129 GB | 128 GB | MIT |
| 31 | GPT-oss 120B 117B MoE · MXFP4 · OpenAI ollama run gpt-oss:120bHugging Face |
68 |
community92communityprefill 2,022 @4KcommunityTTFT 1.109 s @1K | estimated65 GB | 62 GB | Apache 2.0 |
| 32 | Inkling 2 Small 594B MoE · Q4_K_M · Thinking Machines |
68 |
estimated36 M5 Ultra | estimated231 GB | 228 GB | Apache 2.0 |
| 33 | Kolibri-1 78.1B MoE · Q4_K_M · Aleph Alpha |
68 |
estimated75 | estimated52 GB | 50 GB | Apache 2.0 |
| 34 | Ling-3.0 Flash 124B MoE · 4-bit · Ant Group |
68 |
community68communityprefill 1,615 @4KcommunityTTFT 0.644 s @1K | estimated97 GB | 96 GB | MIT |
| 35 | Nanbeige4.2-3B 3B · Q4_K_M · Nanbeige |
68 |
estimated199 | estimated8 GB | 2 GB | Apache 2.0 |
| 36 | Spark-X2.6-4B 4.11B · oQ5e · SparkLLM (XHToken) |
68 |
community102communityprefill 2,867 @8KcommunityTTFT 0.237 s @1K | estimated7 GB | 4 GB | Apache 2.0 |
| 37 | GPT-oss 20B 21B MoE · MXFP4 · OpenAI ollama run gpt-oss:20bHugging Face |
67 |
community141communityprefill 4,774 @4KcommunityTTFT 0.587 s @1K | estimated15 GB | 13 GB | Apache 2.0 |
| 38 | Inkling-Small 276B MoE · Q2_K · Thinking Machines |
67 |
estimated42 | estimated98 GB | 96 GB | Apache 2.0 |
| 39 | Kimi K2.6 1T MoE · 3.6 bpw · Moonshot AI |
67 |
estimated27 M5 Ultra | estimated434 GB | 429 GB | Modified MIT |
| 40 | Kimi K2.7 Code 1T MoE · IQ3_S · Moonshot AI |
67 |
estimated27 M5 Ultra | estimated426 GB | 421 GB | Modified MIT |
| 41 | Ling-3.1-flash 560B MoE · Q4_K_M · Ant Group |
67 |
estimated27 M5 Ultra | estimated361 GB | 360 GB | MIT |
| 42 | Muse Glimmer 30B 30B · Q4_K_M · Meta |
67 |
sourced27sourcedTTFT 0.4 s | estimated19 GB | 18 GB | Apache 2.0 |
| 43 | Qwen3.8-Flash-Next 125B MoE · 4-bit · Alibaba |
67 |
community47communityprefill 1,309 @1K | estimated122 GB | 120 GB | Qwen Custom |
| 44 | GLM-5.3 753B MoE · Q4_K_M · Zhipu AI |
66 |
estimated17 M5 Ultra | estimated262 GB | 256 GB | GLM-5.3 |
| 45 | Gemma 4 12B 12B · Q4_K_M · Google ollama run gemma4:12bHugging Face |
66 |
estimated64 | estimated11 GB | 9 GB | Apache 2.0 |
| 46 | Kimi K2.5 1T MoE · 3.5 bpw · Moonshot AI |
66 |
estimated28 M5 Ultra | estimated422 GB | 417 GB | Modified MIT |
| 47 | MiMo-V2.6-Distill-Qwen-9B 9.41B · Q4_K_M · Xiaomi ollama run mimo-v2.6:9bHugging Face |
66 |
estimated79 | estimated11 GB | 8 GB | MIT |
| 48 | MiniMax M2.5 230B MoE · Q4_K_M · MiniMax |
66 |
estimated57 M5 Ultra | estimated146 GB | 130 GB | Modified MIT |
| 49 | Beam 1.0 501.3B MoE · Q4_K_M · Reflection AI |
65 |
estimated27 M5 Ultra | estimated202 GB | 199 GB | Apache 2.0 |
| 50 | GLM-5.1 754B MoE · Q4_K_M · Zhipu AI |
65 |
estimated17 M5 Ultra | estimated396 GB | 390 GB | MIT |
| 51 | Ling-3.0-tiny 7.9B MoE · oQ4e · Ant Group |
65 |
community174communityprefill 2,012 @4KcommunityTTFT 2.035 s @4K | estimated7 GB | 6 GB | MIT |
| 52 | MiniMax-M3 428B MoE · Q4_K_M · MiniMax |
65 |
estimated42 M5 Ultra | estimated170 GB | 162 GB | MiniMax Community |
| 53 | Muse Spark Open 816B MoE · oQ3e · Meta |
65 |
estimated31 M5 Ultra | — | 342 GB | Muse Model License |
| 54 | Nemotron 3.5 Bolt E4B 4B · Q4_K_M · NVIDIA ollama run nemotron-3.5-bolt:e4bHugging Face |
65 |
estimated161 | estimated8 GB | 6 GB | OpenMDW |
| 55 | Qwen 3 30B-A3B 30B MoE · 4-bit · Alibaba ollama run qwen3:30b-a3bHugging Face |
65 |
sourced152 | estimated25 GB | 18 GB | Apache 2.0 |
| 56 | Qwen 3.5 397B-A17B 397B MoE · 4-bit · Alibaba |
65 |
community62 M5 Ultracommunityprefill 1,265 @8K | estimated212 GB | 210 GB | Apache 2.0 |
| 57 | DeepSeek R1 8B 8B · Q4_K_M · DeepSeek ollama run deepseek-r1:8bHugging Face |
64 |
estimated91 | estimated16 GB | 6 GB | MIT |
| 58 | Gemma 4 31B 31B · 4-bit · Google ollama run gemma4:31bHugging Face |
64 |
community26communityprefill 710 @4KcommunityTTFT 1.439 s @1K | estimated30 GB | 24 GB | Apache 2.0 |
| 59 | Gemma 4 E4B 4B · Q4_K_M · Google ollama run gemma4:e4bHugging Face |
64 |
estimated161 | estimated8 GB | 6 GB | Apache 2.0 |
| 60 | Laguna S 2.1 118B MoE · Q4_K_M · Poolside |
64 |
estimated38 | estimated100 GB | 96 GB | OpenMDW |
| 61 | MiniMax-M3.1-Flash 428B MoE · oQ3e · MiniMax |
64 |
community35 M5 Ultracommunityprefill 1,110 @4KcommunityTTFT 0.969 s @1K | estimated197 GB | 189 GB | MiniMax Community |
| 62 | Phi-4 Mini 3.8B · Q4_K_M · Microsoft ollama run phi4-mini:3.8bHugging Face |
64 |
estimated168 | estimated12 GB | 3 GB | MIT |
| 63 | Step-3.7-Flash 198B MoE · oQ2e · StepFun |
64 |
community48communityprefill 594 @4KcommunityTTFT 1.704 s @1K | estimated69 GB | 65 GB | Apache 2.0 |
| 64 | DeepSeek V4 Pro 1.6T MoE · DeepSeek |
63 |
— | estimated851 GB | 850 GB | MIT |
| 65 | Hy4 preview 770B MoE · Q4_K_M · Tencent |
63 |
estimated14 M5 Ultra | estimated298 GB | 292 GB | Apache 2.0 |
| 66 | LFM3-1.7B 1.73B · Q4_K_M · Liquid AI ollama run lfm3:1.7bHugging Face |
63 |
estimated285 | estimated4 GB | 3 GB | LFM Open |
| 67 | Mellum2.1-12B-A2.5B-Thinking 12.15B MoE · Q4_K_M · JetBrains |
63 |
estimated154 | estimated11 GB | 10 GB | Apache 2.0 |
| 68 | MiMo-V2.6-Pro 1.02T MoE · Xiaomi |
63 |
— | estimated585 GB | 581 GB | MIT |
| 69 | Qwen3-235B-A22B 235B MoE · 4.5 bpw · Alibaba ollama run qwen3:235b-a22bHugging Face |
63 |
estimated26 | estimated140 GB | 128 GB | Apache 2.0 |
| 70 | Apertus 1.5 8B 8B · Q4_K_M · Swiss AI |
62 |
estimated91 | — | 5 GB | Apache 2.0 |
| 71 | DeepSeek V3.2-Speciale 685B MoE · Q4_K_M · DeepSeek |
62 |
estimated18 M5 Ultra | estimated365 GB | 360 GB | MIT |
| 72 | Nemotron 3.5 Super VL 120B-A12B 120B MoE · Q4_0 · NVIDIA ollama run nemotron-3.5-super:120bHugging Face |
62 |
community36communityprefill 885 @4K | estimated73 GB | 72 GB | OpenMDW |
| 73 | dots3-note preview 280B MoE · 3.5 bpw · Xiaohongshu (RedNote) |
62 |
estimated28 | estimated121 GB | 120 GB | Apache 2.0 |
| 74 | Gemma 4 E2B 2B · Q4_K_M · Google ollama run gemma4:e2bHugging Face |
61 |
estimated261 | estimated3 GB | 2 GB | Apache 2.0 |
| 75 | Mistral 7B 7B · Q4_K_M · Mistral ollama run mistral:7bHugging Face |
61 |
estimated102 | 32K max | 5 GB | Apache 2.0 |
| 76 | Qwen 3 8B 8B · Q4_K_M · Alibaba ollama run qwen3:8bHugging Face |
61 |
estimated91 | estimated16 GB | 6 GB | Apache 2.0 |
| 77 | Qwen 3.5 9B 9B · Q4_K_M · Alibaba ollama run qwen3.5:9bHugging Face |
61 |
estimated81 | estimated10 GB | 7 GB | Apache 2.0 |
| 78 | Granite 4.2 8B 8.79B · Q4_K_M · IBM ollama run granite4.2:8bHugging Face |
60 |
estimated84 | estimated19 GB | 8 GB | Apache 2.0 |
| 79 | Hunyuan Hy3 295B MoE · Q4_K_M · Tencent |
60 |
estimated30 M5 Ultra | estimated181 GB | 160 GB | Apache 2.0 |
| 80 | LFM2.5-2.6B 2.6B · Q4_K_M · Liquid AI |
60 |
sourced220sourcedTTFT 0.1 s | estimated4 GB | 2 GB | LFM Open |
| 81 | Ministral 8B 8B · Q4_K_M · Mistral |
60 |
estimated91 | estimated12 GB | 6 GB | Apache 2.0 |
| 82 | Qwen 3.5 122B-A10B 122B MoE · Q4_K_M · Alibaba ollama run qwen3.5:122b-a10bHugging Face |
60 |
community43 | estimated67 GB | 65 GB | Apache 2.0 |
| 83 | Qwen 3.5 4B 4B · Q4_K_M · Alibaba ollama run qwen3.5:4bHugging Face |
60 |
estimated161 | estimated6 GB | 3 GB | Apache 2.0 |
| 84 | DeepSeek V3.2 685B MoE · Q4_K_M · DeepSeek |
59 |
estimated18 M5 Ultra | estimated365 GB | 360 GB | MIT |
| 85 | Mistral Small 4 119B MoE · Q4_K_M · Mistral |
59 |
estimated46 | estimated66 GB | 64 GB | Apache 2.0 |
| 86 | Qwen 3 4B 4B · Q4_K_M · Alibaba ollama run qwen3:4bHugging Face |
59 |
estimated161 | estimated13 GB | 3 GB | Apache 2.0 |
| 87 | Ripple-65B Preview 64.8B · Q4_K_M · Ricursive Intelligence ollama run ripple-preview:65bHugging Face |
59 |
estimated13 | estimated38 GB | 33 GB | Apache 2.0 |
| 88 | Inkling 975B MoE · Thinking Machines |
58 |
— | estimated553 GB | 550 GB | Apache 2.0 |
| 89 | Kimi K3 2.8T MoE · Moonshot AI |
58 |
— | estimated1562 GB | 1560 GB | Moonshot Custom |
| 90 | Kimi K3.1 Pro 2.8T MoE · Moonshot AI |
58 |
— | estimated1565 GB | 1563 GB | Moonshot Custom |
| 91 | MiniMax-M2.7 228.7B MoE · IQ3_XXS · MiniMax |
58 |
estimated45 | estimated98 GB | 82 GB | MiniMax Non-Commercial |
| 92 | Mistral Large 4 1.05T MoE · 3.0 bpw · Mistral |
58 |
estimated21 M5 Ultra | estimated388 GB | 383 GB | Mistral Ecosystem Open-Weight |
| 93 | Qwen3.8-2.4T-A95B 2.4T MoE · Alibaba |
58 |
— | estimated1316 GB | 1310 GB | Qwen Custom |
| 94 | SmolLM3 3B 3B · Q4_K_M · HuggingFace |
58 |
estimated199 | estimated7 GB | 2 GB | Apache 2.0 |
| 95 | Ternary Bonsai 2 9B 9.41B · PQ2_0 · Prism ML ollama run bonsai-2:9bHugging Face |
58 |
sourced77 M5 Pro | estimated8 GB | 5 GB | Apache 2.0 |
| 96 | Llama 3.1 8B 8B · 4-bit · Meta ollama run llama3.1:8bHugging Face |
57 |
sourced102 | — | 5 GB | Meta Custom |
| 97 | Step 5 Preview 600B MoE · StepFun |
57 |
— | estimated266 GB | 261 GB | StepFun Model License |
| 98 | Gemma 3 4B 4B · Q4_K_M · Google ollama run gemma3:4bHugging Face |
56 |
estimated161 | — | 3 GB | Gemma |
| 99 | Granite 4.2 30B 29.28B · Q4_K_M · IBM ollama run granite4.2:30bHugging Face |
56 |
estimated28 | estimated37 GB | 20 GB | Apache 2.0 |
| 100 | Nemotron 3 Ultra 550B-A55B 550B MoE · Q4_K_M · NVIDIA |
56 |
estimated12 M5 Ultra | estimated227 GB | 226 GB | OpenMDW |
| 101 | GLM-4.5-Air 106B MoE · Q4_K_M · Zhipu AI |
55 |
estimated27 | estimated68 GB | 56 GB | MIT |
| 102 | MiMo-V2-Flash 309B MoE · Q4_K_M · Xiaomi |
55 |
estimated41 M5 Ultra | estimated167 GB | 165 GB | MIT |
| 103 | Phi-4 14B 14B · Q4_K_M · Microsoft ollama run phi4:14bHugging Face |
55 |
estimated55 | estimated23 GB | 10 GB | MIT |
| 104 | DeepSeek R1 671B MoE · Q4_K_M · DeepSeek ollama run deepseek-r1:671bHugging Face |
54 |
estimated18 M5 Ultra | estimated355 GB | 350 GB | MIT |
| 105 | DeepSeek V3 685B MoE · Q4_K_M · DeepSeek ollama run deepseek-v3:671bHugging Face |
54 |
estimated18 M5 Ultra | estimated365 GB | 360 GB | MIT |
| 106 | Qwen 3 14B 14B · Q4_K_M · Alibaba ollama run qwen3:14bHugging Face |
54 |
estimated55 | estimated21 GB | 10 GB | Apache 2.0 |
| 107 | Llama 4 Maverick 400B MoE · Q4_K_M · Meta ollama run llama4:maverickHugging Face |
53 |
estimated36 M5 Ultra | — | 210 GB | Meta Custom |
| 108 | Mistral Medium 3.5 128B · Q4_K_M · Mistral ollama run mistral-medium-3.5:128bHugging Face |
53 |
estimated7 | estimated100 GB | 77 GB | Mistral Modified MIT |
| 109 | Mistral Small 3.2 24B 24B · Q4_K_M · Mistral ollama run mistral-small3.2:24bHugging Face |
53 |
estimated33 | estimated25 GB | 14 GB | Apache 2.0 |
| 110 | Devstral Small 24B 24B · Q4_K_M · Mistral |
52 |
estimated33 | estimated25 GB | 14 GB | Apache 2.0 |
| 111 | GLM-4.7 355B MoE · Q4_K_M · Zhipu AI |
52 |
estimated21 M5 Ultra | estimated204 GB | 180 GB | MIT |
| 112 | Ministral 3 14B 14B · Q4_K_M · Mistral ollama run ministral-3:14bHugging Face |
52 |
estimated55 | estimated21 GB | 10 GB | Apache 2.0 |
| 113 | Qwen 2.5 14B 14B · Q4_K_M · Alibaba ollama run qwen2.5:14bHugging Face |
52 |
estimated55 | estimated23 GB | 10 GB | Apache 2.0 |
| 114 | Gemma 3 12B 12B · Q4_K_M · Google ollama run gemma3:12bHugging Face |
51 |
estimated64 | — | 9 GB | Gemma |
| 115 | Mistral Large 3 675B MoE · Q4_K_M · Mistral |
51 |
estimated16 M5 Ultra | estimated360 GB | 355 GB | Apache 2.0 |
| 116 | Solar Open 2 250B 250B MoE · Upstage |
51 |
— | estimated164 GB | 160 GB | Solar License |
| 117 | Step-3.5-Flash 196B MoE · Q4_K_M · StepFun |
51 |
estimated29 | estimated124 GB | 120 GB | Apache 2.0 |
| 118 | DeepSeek R1 32B 32B · Q4_K_M · DeepSeek ollama run deepseek-r1:32bHugging Face |
50 |
estimated25 | estimated39 GB | 22 GB | MIT |
| 119 | QwQ 32B 32B · Q4_K_M · Alibaba ollama run qwq:32bHugging Face |
50 |
estimated25 | 32K max | 22 GB | Apache 2.0 |
| 120 | DeepSeek R1 70B 70B · Q4_K_M · DeepSeek ollama run deepseek-r1:70bHugging Face |
49 |
estimated12 | estimated65 GB | 44 GB | MIT |
| 121 | Qwen 3 32B 32B · 4-bit · Alibaba ollama run qwen3:32bHugging Face |
49 |
sourced26 | estimated39 GB | 22 GB | Apache 2.0 |
| 122 | Qwen 3.5 27B 27B · Q4_K_M · Alibaba ollama run qwen3.5:27bHugging Face |
47 |
estimated30 | estimated23 GB | 18 GB | Apache 2.0 |
| 123 | Apertus 1.5 70B 70B · Q4_K_M · Swiss AI |
46 |
estimated12 | — | 44 GB | Apache 2.0 |
| 124 | Gemma 3 27B 27B · Q4_K_M · Google ollama run gemma3:27bHugging Face |
46 |
estimated30 | — | 18 GB | Gemma |
| 125 | Llama 4 Scout 109B MoE · Q4_K_M · Meta ollama run llama4:scoutHugging Face |
45 |
estimated19 | — | 58 GB | Meta Custom |
| 126 | Qwen 2.5 72B 72B · Q4_K_M · Alibaba ollama run qwen2.5:72bHugging Face |
45 |
estimated12 | estimated67 GB | 46 GB | Apache 2.0 |
| 127 | Llama 3.1 405B 405B · Q4_K_M · Meta ollama run llama3.1:405bHugging Face |
44 |
estimated4 M5 Ultra | — | 240 GB | Meta Custom |
| 128 | Hermes 4 70B 70B · Q4_K_M · Nous Research |
42 |
estimated12 | estimated59 GB | 38 GB | Meta Custom |
| 129 | Llama 3.3 70B 70B · 4-bit · Meta ollama run llama3.3:70bHugging Face |
41 |
sourced13sourcedprefill 306 @0.5K | — | 44 GB | Meta Custom |
| 130 | Mixtral 8x22B 141B MoE · Q4_K_M · Mistral ollama run mixtral:8x22bHugging Face |
41 |
estimated9 | estimated103 GB | 88 GB | Apache 2.0 |
ollama run qwen3.6:35b-a3bHugging Faceollama run glm-4.7-flashHugging Faceollama run flint:40bHugging Faceollama run nemotron-3.5-lightning:30bHugging Faceollama run north-mini-code-1.0Hugging Faceollama run gemma4:26b-a4bHugging Faceollama run openbmb/minicpm-v4.7:35b-a3bHugging Faceollama run lfm2.5:24bHugging Faceollama run qwen3.8:27bHugging Faceollama run nemotron-cascade-2:30bHugging Faceollama run qwen3.5:35b-a3bHugging Faceollama run qwen3.6:27bHugging Faceollama run qwen3-coder-nextHugging Faceollama run bonsai-2:27bHugging Faceollama run gpt-oss:120bHugging Faceollama run gpt-oss:20bHugging Faceollama run gemma4:12bHugging Faceollama run mimo-v2.6:9bHugging Faceollama run nemotron-3.5-bolt:e4bHugging Faceollama run qwen3:30b-a3bHugging Faceollama run deepseek-r1:8bHugging Faceollama run gemma4:31bHugging Faceollama run gemma4:e4bHugging Faceollama run phi4-mini:3.8bHugging Faceollama run lfm3:1.7bHugging Faceollama run qwen3:235b-a22bHugging Faceollama run nemotron-3.5-super:120bHugging Faceollama run gemma4:e2bHugging Faceollama run mistral:7bHugging Faceollama run qwen3:8bHugging Faceollama run qwen3.5:9bHugging Faceollama run granite4.2:8bHugging Faceollama run qwen3.5:122b-a10bHugging Faceollama run qwen3.5:4bHugging Faceollama run qwen3:4bHugging Faceollama run ripple-preview:65bHugging Faceollama run bonsai-2:9bHugging Faceollama run llama3.1:8bHugging Faceollama run gemma3:4bHugging Faceollama run granite4.2:30bHugging Faceollama run phi4:14bHugging Faceollama run deepseek-r1:671bHugging Faceollama run deepseek-v3:671bHugging Faceollama run qwen3:14bHugging Faceollama run llama4:maverickHugging Faceollama run mistral-medium-3.5:128bHugging Faceollama run mistral-small3.2:24bHugging Faceollama run ministral-3:14bHugging Faceollama run qwen2.5:14bHugging Faceollama run gemma3:12bHugging Faceollama run deepseek-r1:32bHugging Faceollama run qwq:32bHugging Faceollama run deepseek-r1:70bHugging Faceollama run qwen3:32bHugging Faceollama run qwen3.5:27bHugging Faceollama run gemma3:27bHugging Faceollama run llama4:scoutHugging Faceollama run qwen2.5:72bHugging Faceollama run llama3.1:405bHugging Faceollama run llama3.3:70bHugging Faceollama run mixtral:8x22bHugging FaceOwn the Mac already? See which of these 130 models fit your exact chip and unified memory →
Still choosing hardware? Three questions and the Mac Advisor names the config → · Every Mac compared for local AI → · Key numbers →
Questions about the ranking
What is the best local LLM for a Mac right now?
According to the LLMCheck index, as of 18 November 2026, MiMo-V2.6-Flash and Qwen 3.6-35B-A3B tie for #1 at LLMCheck Score 81; MiMo-V2.6-Flash is listed first on the alphabetical tie-break. MiMo-V2.6-Flash runs at 77 tok/s community-measured on an M5 Ultra but needs a Mac Studio with at least 192 GB of unified memory. Qwen 3.6-35B-A3B needs only 24 GB and is the best pick on a Mac with 32 GB or less, at 59 tok/s on a Mac mini M6 32 GB (community-measured on an M6). The full ranking of all 130 models is at llmcheck.net/leaderboard.
What is the best local LLM for each amount of RAM?
According to the LLMCheck index, as of 18 November 2026, the best pick for each RAM tier is the best-ranked model that fits the tier, reaches 15 tok/s on the tier's representative Mac and has a standard Mac runtime (the /best-llm/by-ram/ pick), ranking capability against speed: 8 GB Mac: MiniCPM5-2B (Score 69; ~39 tok/s estimated on a MacBook Air M1 8 GB; needs 3 GB); 16 GB Mac: Gemma 4 12B (Score 66; ~17 tok/s estimated on a Mac mini M6 16 GB; needs 9 GB); 24 GB Mac: Recursive Flint (Score 80; ~71 tok/s estimated on a Mac mini M6 24 GB; needs 20 GB); 32 GB Mac: Qwen 3.6-35B-A3B (Score 81; 59 tok/s on a Mac mini M6 32 GB (community-measured on an M6); needs 24 GB); 64 GB Mac: LFM2.5-24B-A2B (Score 76; 114 tok/s on a Mac mini M5 Pro 64 GB (community-measured on an M5 Pro); needs 17 GB); 128 GB Mac: Qwen 3.6-35B-A3B (Score 81; 130 tok/s on a Mac Studio M5 Max 128 GB (community-measured on an M5 Max); needs 24 GB); 192–256 GB Mac: MiMo-V2.6-Flash (Score 81; 77 tok/s on a Mac Studio M5 Ultra 256 GB (community-measured on an M5 Ultra); needs 170 GB); 512 GB Mac: MiMo-V2.6-Flash (Score 81; 77 tok/s on a Mac Studio M5 Ultra 512 GB (community-measured on an M5 Ultra); needs 170 GB). Each tier, with every Mac in it: llmcheck.net/best-llm/by-ram/.
What is the best dense local LLM?
The strongest dense (non-mixture-of-experts) model in the index is Qwen3.8-27B (capability 46/50, 19 GB minimum, 31 tok/s community-measured on an M5 Max at 4-bit). Dense models read every weight for every token, so they decode slower than mixture-of-experts models of similar size.
What is the fastest local LLM on a Mac?
The fastest measured figure in the index is Maple Preview 20B-A1B at 387 tok/s on an M5 Pro (published, Ternary); its ternary weights decode through compute-bound kernels, so among bandwidth-bound models the fastest measured reference figure is LFM2.5-2.6B at 220 tok/s published on an M5 Max (Q4_K_M).
How is the LLMCheck Score calculated?
Capability (0-50, normalized from published benchmarks) + speed (0-25: the displayed tok/s x 0.25, capped) + RAM accessibility (0-15: 15 at 8 GB or less, down to 3 above 64 GB) + licence openness (0-10: MIT 10 down to non-commercial 2). The speed figure is the lower-median measured plain-decode run on an M5 Max (or an M5 Ultra for models above 128 GB) where one exists, otherwise the published bandwidth estimate. Method: llmcheck.net/methodology#reference-row.
Why did models move in the 19 November 2026 update?
The 19 November 2026 update (data of 18 November) changed no ranking rule. It added a RAM-at-64K figure for each model whose KV-cache layout has been read from its config.json (minimum RAM covers about the first 4K tokens of context; a longer context adds its cache), decode-at-depth and prefill figures where a run measured them, and speculative-decoding (MTP) figures kept beside their paired plain run. None of these enters the Score. As of 18 November 2026, 320 measured rows cover 61 of the 130 models (all listed at llmcheck.net/benchmarks#measured); for 35 of the 130 models the leaderboard figure itself is measured, and the rest show an estimate on their reference chip. The largest gaps between a measurement and the formula's own estimate at the same quant: Nemotron 3.5 Bolt 74B-A9B 64 tok/s community-measured vs 30 estimated on an M5 Max (5-bit); MiMo-V2.6-Flash 77 tok/s community-measured vs 42 estimated on an M5 Ultra (MXFP4); Qwen 3.5 397B-A17B 62 tok/s community-measured vs 36 estimated on an M5 Ultra (4-bit). GLM-4.7-Flash's M5 Max figure (~107 tok/s, estimated) is scaled from its own community-measured M6 run of 46.6 tok/s. Every change is logged at llmcheck.net/changelog.
How much memory does a 64K context need?
According to the LLMCheck index, a model's RAM at 64K is its minimum RAM, which covers about the first 4K tokens, plus the 16-bit KV cache the rest of a 65,536-token context adds, computed from the attention layout in its config.json: an estimate, not a measurement. Qwen 3.6-35B-A3B needs 24 GB at a short context and about 26 GB at 64K (25 GB with an 8-bit KV cache). Sliding-window and linear-attention layers add less. Of the 117 with a recorded layout, 51 are agent-ready at 64K on a 32 GB Mac. Method: llmcheck.net/methodology#context.
LLMCheck. (2026). The local LLM leaderboard. llmcheck.net. https://llmcheck.net/leaderboard (updated ). CC BY 4.0.