| Model | Size | At 4-bit | Smallest Mac |
|---|---|---|---|
| Qwen3 0.6BQwen (Alibaba) | 0.6B | 1.8 GB | 8 GB |
| Qwen3.5 0.8BQwen (Alibaba) | 0.8B | 2 GB | 8 GB |
| Llama 3.2 1B InstructMeta | 1.23B | 2.2 GB | 8 GB |
| Qwen3.5 2BQwen (Alibaba) | 2B | 2.6 GB | 8 GB |
| MiniCPM5 2BOpenBMB | 2.52B | 2.9 GB | 8 GB |
| LFM2.5 2.6BLiquid AI | 2.69B | 3 GB | 8 GB |
| Granite 4.2 3BIBM | 3B | 3.2 GB | 8 GB |
| SmolLM3 3BHugging Face | 3B | 3.2 GB | 8 GB |
| Llama 3.2 3B InstructMeta | 3.21B | 3.3 GB | 8 GB |
| Phi-4 mini instructMicrosoft | 3.8B | 3.6 GB | 8 GB |
| Nemotron 3 Nano 4BNVIDIA | 3.97B | 3.7 GB | 8 GB |
| Qwen3.5 4BQwen (Alibaba) | 4B | 3.8 GB | 8 GB |
| Gemma 4 E2BGoogle DeepMind | 5.1B | 4.4 GB | 8 GB |
| Llama 3.1 8B InstructMeta | 8B | 6 GB | 8 GB |
| Gemma 4 E4BGoogle DeepMind | 8B | 6 GB | 8 GB |
| Granite 4.2 8BIBM | 8B | 6 GB | 8 GB |
| Qwen3 8BQwen (Alibaba) | 8.2B | 6.1 GB | 16 GB |
| Ministral 3 8BMistral AI | 8.9B | 6.5 GB | 16 GB |
| Qwen3.5 9BQwen (Alibaba) | 9B | 6.6 GB | 16 GB |
| Gemma 4 12B UnifiedGoogle DeepMind | 12B | 8.2 GB | 16 GB |
| Phi-4 (14B)Microsoft | 14B | 9.4 GB | 16 GB |
| Phi-4 Reasoning Vision 15BMicrosoft | 15B | 9.9 GB | 16 GB |
| gpt-oss-20bOpenAI | 21B · 3.6B active | 15.3 GB | 24 GB |
| Devstral Small 2 (24B)Mistral AI | 24B | 15 GB | 24 GB |
| Magistral Small 1.2 (2509)Mistral AI | 24B | 15 GB | 24 GB |
| Mistral Small 3.2 (24B)Mistral AI | 24B | 15 GB | 24 GB |
| LFM2.5 24B A2BLiquid AI | 24B · 2B active | 15 GB | 24 GB |
| Gemma 4 26B A4BGoogle DeepMind | 25.2B · 3.8B active | 15.7 GB | 24 GB |
| Qwen3.8 27BQwen (Alibaba) | 27B | 16.7 GB | 24 GB |
| Qwen3.6 27BQwen (Alibaba) | 27B | 16.7 GB | 24 GB |
| Muse Glimmer 30BMeta (Meta Superintelligence Lab) | 29.6B | 18.2 GB | 32 GB |
| GLM-4.7 FlashZ.ai (Zhipu) | 30B · 3B active | 18.4 GB | 32 GB |
| Nemotron 3 Nano 30B A3BNVIDIA | 30B · 3.5B active | 18.4 GB | 32 GB |
| Nemotron 3.5 Lightning 30B A3BNVIDIA | 30B · 3B active | 18.4 GB | 32 GB |
| Granite 4.2 30BIBM | 30B | 18.4 GB | 32 GB |
| Qwen3 Coder 30B A3B InstructQwen (Alibaba) | 30.5B · 3.3B active | 18.7 GB | 32 GB |
| Gemma 4 31BGoogle DeepMind | 30.7B | 18.8 GB | 32 GB |
| Qwen3.6 35B A3BQwen (Alibaba) | 35B · 3B active | 21.3 GB | 32 GB |
| Llama 3.3 70B InstructMeta | 70B | 42.6 GB | 64 GB |
| Nemotron 3.5 Bolt 74B A9BNVIDIA | 74B · 9B active | 44.9 GB | 64 GB |
| Qwen3 Coder NextQwen (Alibaba) | 80B · 3B active | 48.6 GB | 96 GB |
| GLM-4.5 AirZ.ai (Zhipu) | 106B · 12B active | 64.4 GB | 96 GB |
| Llama 4 Scout 17B-16E InstructMeta | 109B · 17B active | 66.2 GB | 96 GB |
| gpt-oss-120bOpenAI | 117B · 5.1B active | 70.5 GB | 96 GB |
| Mistral Small 4 (119B)Mistral AI | 119B · 6.5B active | 72.3 GB | 128 GB |
| Nemotron 3 Super 120B A12BNVIDIA | 120B · 12B active | 72.9 GB | 128 GB |
| Qwen3.5 122B A10BQwen (Alibaba) | 122B · 10B active | 74.1 GB | 128 GB |
| Ornith 1.5 122BOrnith AI | 122B · 10B active | 74.1 GB | 128 GB |
| Devstral 2 (123B)Mistral AI | 123B | 74.7 GB | 128 GB |
| Mistral Medium 3.5 (128B)Mistral AI | 128B | 77.8 GB | 128 GB |
| Qwen3.8 Flash NextQwen (Alibaba) | 180B · 6B active | 109 GB | 192 GB |
| MiniMax M2.7MiniMax | 230B · 10B active | 140 GB | 192 GB |
| Inkling SmallThinking Machines Lab | 276B · 12B active | 168 GB | 256 GB |
| DeepSeek V4 Flash 0731DeepSeek | 284B · 13B active | 173 GB | 256 GB |
| Hy3Tencent | 295B · 21B active | 179 GB | 256 GB |
| MiMo V2.6 FlashXiaomi | 309B · 15B active | 187 GB | 256 GB |
| GLM-5.3 FlashZ.ai (Zhipu) | 320B · 18B active | 194 GB | 512 GB |
| Qwen3.5 397B A17BQwen (Alibaba) | 397B · 17B active | 241 GB | 512 GB |
| Llama 4 Maverick 17B-128E InstructMeta | 400B · 17B active | 243 GB | 512 GB |
| MiniMax M3MiniMax | 428B · 23B active | 260 GB | 512 GB |
| Beam 1.0Reflection AI | 501B · 23B active | 304 GB | 512 GB |
| Nemotron 3 Ultra 550B A55BNVIDIA | 550B · 55B active | 334 GB | 512 GB |
| DeepSeek V4.1 Flash 1017DeepSeek | 552B · 16B active | 335 GB | 512 GB |
| Mistral Large 3 (675B)Mistral AI | 675B · 41B active | 410 GB | 512 GB |
| GLM-5.3Z.ai (Zhipu) | 744B · 40B active | 452 GB | 512 GB |
| Muse Spark OpenMeta (Meta Superintelligence Labs) | 816B · 42B active | 496 GB | 512 GB |
| InklingThinking Machines Lab | 975B · 41B active | 592 GB | 512 GB |
| Kimi K2.6Moonshot AI | 1,000B · 32B active | 643 GB | 512 GB |
| Kimi K2.7 CodeMoonshot AI | 1,000B · 32B active | 643 GB | 512 GB |
| MiMo V2.6 ProXiaomi | 1,020B · 42B active | 611 GB | 512 GB |
| DeepSeek V4 Pro 0813DeepSeek | 1,600B · 49B active | 972 GB | none |
| Qwen3.8 2.4T A95BQwen (Alibaba) | 2,400B · 95B active | 1.5 TB | none |
| Kimi K3Moonshot AI | 2,800B · 104B active | 1.7 TB | none |
Or start from your Mac: every Mac · compare
How it’s worked out
Memory: the model’s parameters at 8, 6, 4, 3 or 2 bits per weight plus MLX’s group scales (half a bit per weight), or its native format, with room for the context on top. By default macOS lets the GPU use about 75% of unified memory. Speed: memory bandwidth divided by the bytes read per token is the ceiling for one conversation; in llama.cpp’s public Apple Silicon benchmarks real runs reach about 45% to 80% of it, which is the range shown. These are estimates, not measurements: measure your own Mac on MLX Bench. Model figures from each official model card, chip figures from Apple’s Tech Specs. Checked 2026-10-19. Same data as mlx.app/compare.