Apple silicon (unified memory)
What AI models can the Apple M3 Ultra run locally?
At 4-bit quantization with an 8K-token conversation, 119 of the 146 open-weight models in our directory fit entirely in the Apple M3 Ultra's 512 GB of unified memory (about 384 GB usable, the default GPU share, often adjustable), up to Step 5 Preview. Among models of 7B parameters or more, Laguna XS 2.1 (free) is the fastest, at an estimated 190 to 245 tokens per second.
| Model | Parameters | Memory needed | Estimated speed |
|---|---|---|---|
| Runs entirely in fast memory (119) | |||
| Step 5 Preview StepFun | 604B | 378 GB 366 GB + 11 GB cache | 0.8–1.0 tokens/s Too slow for chat, little memory to spare |
| Nemotron Ultra NVIDIA | 561B | 341 GB weights only | Speed not estimated (architecture not public) |
| Beam 1.0 501B A23B Reflection | 501B | 305 GB 304 GB + 0.4 GB cache | Speed not estimated for this architecture |
| Qwen3 Coder 480B A35B Alibaba (Qwen) | 480B 35.5B active | 294 GB 291 GB + 2.1 GB cache | 13–17 tokens/s Comfortable to read |
| MiniMax-M3 MiniMax | 427B | 261 GB 259 GB + 1.0 GB cache | Speed not estimated for this architecture |
| ERNIE 4.5 VL 424B A47B Baidu | 424B | 259 GB 257 GB + 1.8 GB cache | 1.1–1.5 tokens/s Too slow for chat |
| Hermes 4 405B Nousresearch | 406B | 251 GB 246 GB + 4.2 GB cache | 1.2–1.5 tokens/s Too slow for chat |
| Hermes 3 405B Instruct Nousresearch | 406B | 251 GB 246 GB + 4.2 GB cache | 1.2–1.5 tokens/s Too slow for chat |
| Hermes 3 405B Instruct Nousresearch | 406B | 251 GB 246 GB + 4.2 GB cache | 1.2–1.5 tokens/s Too slow for chat |
| Qwen3.5 397B-A17B Alibaba (Qwen) | 403B 24.4B active | 246 GB 245 GB + 0.3 GB cache | 20–26 tokens/s Comfortable to read |
| Llama 4 Maverick Meta | 402B | 244 GB weights only | Speed not estimated (architecture not public) |
| Trinity Large Thinking Arcee AI | 399B | 244 GB 242 GB + 1.3 GB cache | Speed not estimated for this architecture |
| Nex-N2.5-Pro NEX AGI | 397B 17.8B active | 242 GB 241 GB + 0.3 GB cache | 27–35 tokens/s Fast |
| GLM-4.5 Zhipu AI (Z.ai) | 358B 39.2B active | 221 GB 217 GB + 3.1 GB cache | 12–15 tokens/s Comfortable to read |
| GLM-4.7 Zhipu AI (Z.ai) | 358B 39.2B active | 221 GB 217 GB + 3.1 GB cache | 12–15 tokens/s Comfortable to read |
| GLM-4.6 Zhipu AI (Z.ai) | 357B 37.6B active | 220 GB 216 GB + 3.1 GB cache | 12–16 tokens/s Comfortable to read |
| GLM-5.3-Flash Zhipu AI (Z.ai) | 321B 25.4B active | 196 GB 195 GB + 0.1 GB cache | 19–25 tokens/s Comfortable to read |
| MiMo-V2.5 Xiaomi (MiMo) | 311B | 190 GB 188 GB + 1.2 GB cache | Speed not estimated for this architecture |
| DeepSeek V4 Flash Vision Exp DeepSeek | 305B 34.1B active | 186 GB 185 GB + 0.7 GB cache | 14–18 tokens/s Comfortable to read |
| DeepSeek V4 Flash 0731 DeepSeek | 304B 33.6B active | 186 GB 184 GB + 0.7 GB cache | 14–18 tokens/s Comfortable to read |
| Hy3 preview Tencent (Hunyuan) | 299B 24.4B active | 185 GB 181 GB + 2.7 GB cache | 18–24 tokens/s Comfortable to read |
| Hy3 Tencent (Hunyuan) | 299B 24.4B active | 185 GB 181 GB + 2.7 GB cache | 18–24 tokens/s Comfortable to read |
| DeepSeek V4 Flash DeepSeek | 291B 20.4B active | 178 GB 176 GB + 0.7 GB cache | 23–30 tokens/s Comfortable to read |
| Inkling Small Thinkingmachines | 266B | 163 GB 161 GB + 1.4 GB cache | Speed not estimated for this architecture |
| Qwen3 VL 235B A22B Instruct Alibaba (Qwen) | 236B 22.8B active | 145 GB 143 GB + 1.6 GB cache | 20–26 tokens/s Comfortable to read |
| Qwen3 VL 235B A22B Thinking Alibaba (Qwen) | 236B 22.8B active | 145 GB 143 GB + 1.6 GB cache | 20–26 tokens/s Comfortable to read |
| Qwen3 235B A22B Instruct 2507 Alibaba (Qwen) | 235B 22.2B active | 145 GB 143 GB + 1.6 GB cache | 21–27 tokens/s Comfortable to read |
| Qwen3 235B A22B Thinking 2507 Alibaba (Qwen) | 235B 22.2B active | 145 GB 143 GB + 1.6 GB cache | 21–27 tokens/s Comfortable to read |
| Qwen3 235B-A22B Alibaba (Qwen) | 235B 22.2B active | 145 GB 143 GB + 1.6 GB cache | 21–27 tokens/s Comfortable to read |
| MiniMax-M2.5 MiniMax | 229B 11.0B active | 142 GB 139 GB + 2.1 GB cache | 38–49 tokens/s Fast |
| MiniMax-M2.7 MiniMax | 229B 11.0B active | 142 GB 139 GB + 2.1 GB cache | 38–49 tokens/s Fast |
| MiniMax-M2 MiniMax | 229B 11.0B active | 142 GB 139 GB + 2.1 GB cache | 38–49 tokens/s Fast |
| MiniMax-M2.1 MiniMax | 229B 11.0B active | 142 GB 139 GB + 2.1 GB cache | 38–49 tokens/s Fast |
| Step 3.7 Flash StepFun | 201B | 127 GB 122 GB + 3.8 GB cache | 2.4–3.1 tokens/s Too slow for chat |
| Step 3.5 Flash StepFun | 199B | 126 GB 121 GB + 3.8 GB cache | 2.4–3.1 tokens/s Too slow for chat |
| Ling 3.0 Flash Inclusionai | 127B 8.6B active | 78 GB 77 GB + 0.4 GB cache | 55–71 tokens/s Fast |
| Ling 3.0 Flash Inclusionai | 127B 8.6B active | 78 GB 77 GB + 0.4 GB cache | 55–71 tokens/s Fast |
| Qwen3.5 122B-A10B Alibaba (Qwen) | 125B 12.7B active | 77 GB 76 GB + 0.2 GB cache | 38–49 tokens/s Fast |
| Ling 3.0 Flash VL Inclusionai | 125B 5.9B active | 77 GB 76 GB + 0.4 GB cache | 78–100 tokens/s Fast |
| Ling 3.0 Flash VL Inclusionai | 125B 5.9B active | 77 GB 76 GB + 0.4 GB cache | 78–100 tokens/s Fast |
| Nemotron Super NVIDIA | 124B | 76 GB 75 GB + 0.1 GB cache | Speed not estimated for this architecture |
| Laguna S 2.1 (free) Poolside | 118B 6.1B active | 73 GB 71 GB + 0.5 GB cache | 75–97 tokens/s Fast |
| OpenAI GPT 120B OpenAI | 117B 5.8B active | 72 GB 71 GB + 0.3 GB cache | 81–105 tokens/s Fast |
| GLM-4.5-Air Zhipu AI (Z.ai) | 110B 17.0B active | 69 GB 67 GB + 1.5 GB cache | 27–34 tokens/s Fast |
| Llama 4 Scout Meta | 109B | 67 GB weights only | Speed not estimated (architecture not public) |
| GLM-4.6V Zhipu AI (Z.ai) | 108B 14.3B active | 68 GB 65 GB + 1.5 GB cache | 31–40 tokens/s Fast |
| GLM-4.5V Zhipu AI (Z.ai) | 108B 14.3B active | 68 GB 65 GB + 1.5 GB cache | 31–40 tokens/s Fast |
| Qwen3-Next 80B-A3B Instruct Alibaba (Qwen) | 81.3B 5.5B active | 51 GB 49 GB + 0.8 GB cache | 79–102 tokens/s Fast |
| Qwen3-Next 80B-A3B (Thinking) Alibaba (Qwen) | 81.3B 5.5B active | 51 GB 49 GB + 0.8 GB cache | 79–102 tokens/s Fast |
| Hunyuan A13B Instruct Tencent (Hunyuan) | 80.4B | 51 GB 49 GB + 1.1 GB cache | Speed not estimated for this architecture |
| Qwen3 Coder Next Alibaba (Qwen) | 79.7B 3.9B active | 50 GB 48 GB + 0.8 GB cache | 107–139 tokens/s Fast |
| Qwen2.5-VL 72B Instruct Alibaba (Qwen) | 73.4B | 48 GB 45 GB + 2.7 GB cache | 6.4–8.3 tokens/s Usable but slow |
| Qwen2.5 72B Instruct Alibaba (Qwen) | 72.7B | 48 GB 44 GB + 2.7 GB cache | 6.5–8.4 tokens/s Usable but slow |
| Magnum v4 72B Anthracite ORG | 72.7B | 48 GB 44 GB + 2.7 GB cache | 6.5–8.4 tokens/s Usable but slow |
| Hermes 3 70B Instruct Nousresearch | 70.6B | 46 GB 43 GB + 2.7 GB cache | 6.7–8.7 tokens/s Usable but slow |
| Hermes 3 70B Instruct Nousresearch | 70.6B | 46 GB 43 GB + 2.7 GB cache | 6.7–8.7 tokens/s Usable but slow |
| Llama-3.3-70B-Instruct Meta | 70.6B | 44 GB weights only | Speed not estimated (architecture not public) |
| Llama-3.1-70B-Instruct Meta | 70.6B | 44 GB weights only | Speed not estimated (architecture not public) |
| Skyfall 36B V2 Thedrummer | 36.9B | 25 GB 22 GB + 2.1 GB cache | 13–16 tokens/s Comfortable to read |
| Qwen3.5 35B-A3B Alibaba (Qwen) | 36.0B 4.7B active | 23 GB 22 GB + 0.2 GB cache | 100–129 tokens/s Fast |
| Qwen3.6 35B-A3B Alibaba (Qwen) | 36.0B 4.7B active | 23 GB 22 GB + 0.2 GB cache | 100–129 tokens/s Fast |
| Nex-N2.5-Mini NEX AGI | 35.1B 3.9B active | 22 GB 21 GB + 0.2 GB cache | 120–156 tokens/s Fast |
| Laguna XS 2.1 (free) Poolside | 33.4B 2.2B active | 21 GB 20 GB + 0.4 GB cache | 190–245 tokens/s Fast |
| Qwen3 VL 32B Instruct Alibaba (Qwen) | 33.4B | 23 GB 20 GB + 2.1 GB cache | 14–18 tokens/s Comfortable to read |
| Qwen2.5 Coder 32B Instruct Alibaba (Qwen) | 32.8B | 23 GB 20 GB + 2.1 GB cache | 14–18 tokens/s Comfortable to read |
| Qwen3 32B Alibaba (Qwen) | 32.8B | 23 GB 20 GB + 2.1 GB cache | 14–18 tokens/s Comfortable to read |
| Nemotron 3 Nano 30B A3B NVIDIA | 31.6B | 20 GB 19 GB + 0.1 GB cache | Speed not estimated for this architecture |
| Nemotron 3.5 Lightning 30B A3B NVIDIA | 31.6B | 20 GB 19 GB + 0.4 GB cache | Speed not estimated for this architecture |
| Gemma 4 31B IT Google DeepMind | 31.3B | 22 GB 19 GB + 2.2 GB cache | 15–19 tokens/s Comfortable to read |
| GLM-4.7-Flash Zhipu AI (Z.ai) | 31.2B 5.2B active | 20 GB 19 GB + 0.4 GB cache | 88–114 tokens/s Fast |
| Qwen3 VL 30B A3B Instruct Alibaba (Qwen) | 31.1B 3.9B active | 20 GB 19 GB + 0.8 GB cache | 107–138 tokens/s Fast |
| Qwen3 VL 30B A3B Thinking Alibaba (Qwen) | 31.1B 3.9B active | 20 GB 19 GB + 0.8 GB cache | 107–138 tokens/s Fast |
| Qwen3-Coder 30B-A3B Instruct Alibaba (Qwen) | 30.5B 3.4B active | 20 GB 19 GB + 0.8 GB cache | 121–157 tokens/s Fast |
| Qwen3 30B A3B Instruct 2507 Alibaba (Qwen) | 30.5B 3.4B active | 20 GB 19 GB + 0.8 GB cache | 121–157 tokens/s Fast |
| Qwen3 30B A3B Thinking 2507 Alibaba (Qwen) | 30.5B 3.4B active | 20 GB 19 GB + 0.8 GB cache | 121–157 tokens/s Fast |
| Qwen3 30B A3B Alibaba (Qwen) | 30.5B 3.4B active | 20 GB 19 GB + 0.8 GB cache | 121–157 tokens/s Fast |
| Hy-MT2-30B-A3B Tencent (Hunyuan) | 30.1B 3.5B active | 20 GB 18 GB + 0.8 GB cache | 118–153 tokens/s Fast |
| Muse Glimmer 30B Meta | 29.8B | 19 GB 18 GB + 0.2 GB cache | 16–21 tokens/s Comfortable to read |
| Qwen3.8 27B Alibaba (Qwen) | 27.8B | 18 GB 17 GB + 0.5 GB cache | 17–22 tokens/s Comfortable to read |
| Qwen3.5 27B Alibaba (Qwen) | 27.8B | 18 GB 17 GB + 0.5 GB cache | 17–22 tokens/s Comfortable to read |
| Qwen3.6 27B Alibaba (Qwen) | 27.8B | 18 GB 17 GB + 0.5 GB cache | 17–22 tokens/s Comfortable to read |
| Gemma 3 27B IT Google DeepMind | 27.4B | 17 GB weights only | Speed not estimated (architecture not public) |
| Gemma 2 27B Google DeepMind | 27.2B | 17 GB weights only | Speed not estimated (architecture not public) |
| Gemma 4 26B A4B IT Google DeepMind | 25.8B | 17 GB 16 GB + 0.5 GB cache | Speed not estimated for this architecture |
| Voxtral Small 24B 2507 Mistral AI | 24.3B | 17 GB 15 GB + 1.3 GB cache | 19–25 tokens/s Comfortable to read |
| Uncensored Cognitivecomputations | 24.0B | 17 GB 15 GB + 1.3 GB cache | 19–25 tokens/s Comfortable to read |
| Mistral Small 3.1 24B Instruct Mistral AI | 24.0B | 17 GB 15 GB + 1.3 GB cache | 19–25 tokens/s Comfortable to read |
| LFM2.5-24B-A2B Liquid | 24.0B | 16 GB 15 GB + 0.2 GB cache | Speed not estimated for this architecture |
| Mistral Small 3 Mistral AI | 23.6B | 16 GB 14 GB + 1.3 GB cache | 20–26 tokens/s Comfortable to read |
| Cydonia 24B V4.1 Thedrummer | 23.6B | 16 GB 14 GB + 1.3 GB cache | 20–26 tokens/s Comfortable to read |
| Cydonia 24B V4.1 Thedrummer | 23.6B | 16 GB 14 GB + 1.3 GB cache | 20–26 tokens/s Comfortable to read |
| Safety GPT OSS 20B OpenAI | 21.5B 4.8B active | 14 GB 13 GB + 0.2 GB cache | 98–127 tokens/s Fast |
| GPT OSS 20B OpenAI | 20.9B 4.2B active | 14 GB 13 GB + 0.2 GB cache | 112–144 tokens/s Fast |
| Reka Flash 3 Rekaai | 20.9B | 15 GB 13 GB + 1.1 GB cache | 22–29 tokens/s Comfortable to read |
| Qwen3 14B Alibaba (Qwen) | 14.8B | 11 GB 9.0 GB + 1.3 GB cache | 31–40 tokens/s Fast |
| Phi 4 Microsoft | 14.7B | 11 GB 8.9 GB + 1.7 GB cache | 30–39 tokens/s Fast |
| Ministral 3 14B Mistral AI | 13.9B | 11 GB 8.5 GB + 1.3 GB cache | 32–42 tokens/s Fast |
| Gemma 3 12B IT Google DeepMind | 12.2B | 8.2 GB weights only | Speed not estimated (architecture not public) |
| Llama Guard 4 12B Meta | 12.0B | 8.1 GB weights only | Speed not estimated (architecture not public) |
| Qwen3.5 9B Alibaba (Qwen) | 9.7B | 6.9 GB 5.9 GB + 0.3 GB cache | 49–64 tokens/s Fast |
| Moondream 3.1 Moondream | 9.3B | 6.4 GB weights only | Speed not estimated (architecture not public) |
| Moondream 3 Preview Moondream | 9.3B | 6.4 GB weights only | Speed not estimated (architecture not public) |
| Ministral 3 8B Mistral AI | 8.9B | 7.3 GB 5.4 GB + 1.1 GB cache | 49–64 tokens/s Fast |
| Granite 4.2 8B IBM | 8.8B | 7.5 GB 5.3 GB + 1.3 GB cache | 49–64 tokens/s Fast |
| Qwen3 VL 8B Instruct Alibaba (Qwen) | 8.8B | 7.3 GB 5.3 GB + 1.2 GB cache | 50–64 tokens/s Fast |
| Qwen3 VL 8B Thinking Alibaba (Qwen) | 8.8B | 7.3 GB 5.3 GB + 1.2 GB cache | 50–64 tokens/s Fast |
| UI-TARS 7B Bytedance | 8.3B | 6.3 GB 5.0 GB + 0.5 GB cache | 56–73 tokens/s Fast |
| Qwen3 8B Alibaba (Qwen) | 8.2B | 7.0 GB 5.0 GB + 1.2 GB cache | 53–69 tokens/s Fast |
| Llama 3.1 8B Instruct fp8 Meta | 8.0B | 5.7 GB weights only | Speed not estimated (architecture not public) |
| Hy-MT2-7B Tencent (Hunyuan) | 8.0B | 6.7 GB 4.9 GB + 1.1 GB cache | 55–71 tokens/s Fast |
| Qwen2.5 7B Instruct Alibaba (Qwen) | 7.6B | 5.9 GB 4.6 GB + 0.5 GB cache | 61–79 tokens/s Fast |
| Gemma 3 4B IT Google DeepMind | 4.3B | 3.4 GB weights only | Speed not estimated (architecture not public) |
| Nemotron 3.5 Content Safety (free) NVIDIA | 4.3B | 3.7 GB 2.6 GB + 0.3 GB cache | 107–139 tokens/s Fast |
| Ministral 3 3B Mistral AI | 3.8B | 4.0 GB 2.3 GB + 0.9 GB cache | 106–138 tokens/s Fast |
| Llama 3.2 3B Instruct Meta | 3.2B | 2.7 GB weights only | Speed not estimated (architecture not public) |
| Granite 4.0 Micro IBM | 3.2B | 2.8 GB 1.9 GB + 0.1 GB cache | 150–194 tokens/s Fast |
| Hy-MT2-1.8B Tencent (Hunyuan) | 2.0B | 2.6 GB 1.2 GB + 0.5 GB cache | 196–254 tokens/s Fast |
| Hy-MT2-1.8B Tencent (Hunyuan) | 2.0B | 2.6 GB 1.2 GB + 0.5 GB cache | 196–254 tokens/s Fast |
| Llama 3.2 1B Instruct Meta | 1.2B | 1.5 GB weights only | Speed not estimated (architecture not public) |
27 larger models do not fit. Change the quantization, conversation length or memory size in the tool.
Compared with similar chips
- Apple M5 Max (40-core GPU): 128 GB, 84 models fit entirely (-35); about 13% slower on models that fit both (614 vs 819 GB/s).
Frequently asked questions
What is the largest AI model the Apple M3 Ultra can run?
At 4-bit quantization with an 8K-token conversation, the largest model in our directory that fits entirely in its 512 GB of unified memory is Step 5 Preview (378 GB).
Can the Apple M3 Ultra run GPT OSS 20B?
Yes. At 4-bit, GPT OSS 20B needs about 14 GB and fits entirely in the Apple M3 Ultra's memory at an estimated 112 to 144 tokens per second.
Can the Apple M3 Ultra run Qwen3 32B?
Yes. At 4-bit, Qwen3 32B needs about 23 GB and fits entirely in the Apple M3 Ultra's memory at an estimated 14 to 18 tokens per second.
Can the Apple M3 Ultra run OpenAI GPT 120B?
Yes. At 4-bit, OpenAI GPT 120B needs about 72 GB and fits entirely in the Apple M3 Ultra's memory at an estimated 81 to 105 tokens per second.
These are estimates, not measurements: memory from each model's parameter count and published architecture, speed from memory bandwidth, calibrated on the AIPCs test bench. Drivers, software and prompt length move real results. How the estimate works.