Apple
Best local LLMs for the Mac Studio · M3 Ultra, 512 GB
512 GB unified memory · 819 GB/s · ~369 GB usable. Enough unified memory to load DeepSeek-R1 671B at Q4 in one machine.
Try it with your context & use case
Preset to the Mac Studio · M3 Ultra, 512 GB. Change the context length or filter by use case.
29 models run well and 1 run tight on Mac Studio · M3 Ultra, 512 GB (any context — just fitting the weights) — 369 GB usable (≈72% of unified memory).
| Model | Size | Fit | Best quant | Needs | Memory | Speed |
|---|---|---|---|---|---|---|
| Qwen3 235B-A22B (MoE)MoE | 235B | Runs well | Q8_0 | 243 GB | ~16 tok/s | |
| Llama 4 Scout (MoE)MoE | 109B | Runs well | FP16 | 212 GB | ~11 tok/s | |
| Qwen2.5 72B | 72.7B | Runs well | FP16 | 142 GB | ~4.1 tok/s | |
| Llama 3.3 70B | 70.6B | Runs well | FP16 | 138 GB | ~4.2 tok/s | |
| DeepSeek-R1 Distill Llama 70B | 70.6B | Runs well | FP16 | 138 GB | ~4.2 tok/s | |
| Mixtral 8x7B (MoE)MoE | 46.7B | Runs well | FP16 | 91 GB | ~15 tok/s | |
| Qwen3 32B | 32.8B | Runs well | FP16 | 64 GB | ~9.0 tok/s | |
| Qwen2.5 32B | 32.8B | Runs well | FP16 | 64 GB | ~9.0 tok/s | |
| Qwen2.5-Coder 32B | 32.8B | Runs well | FP16 | 64 GB | ~9.0 tok/s | |
| DeepSeek-R1 Distill Qwen 32B | 32.8B | Runs well | FP16 | 64 GB | ~9.0 tok/s | |
| Qwen3 30B-A3B (MoE)MoE | 30.5B | Runs well | FP16 | 60 GB | ~58 tok/s | |
| Gemma 3 27B | 27.4B | Runs well | FP16 | 54 GB | ~11 tok/s | |
| Gemma 2 27B | 27.2B | Runs well | FP16 | 53 GB | ~11 tok/s | |
| Mistral Small 3 24B | 23.6B | Runs well | FP16 | 46 GB | ~12 tok/s | |
| Qwen3 14B | 14.8B | Runs well | FP16 | 29 GB | ~20 tok/s | |
| Qwen2.5 14B | 14.8B | Runs well | FP16 | 29 GB | ~20 tok/s | |
| DeepSeek-R1 Distill Qwen 14B | 14.8B | Runs well | FP16 | 29 GB | ~20 tok/s | |
| Phi-4 14B | 14.7B | Runs well | FP16 | 29 GB | ~20 tok/s | |
| Mistral Nemo 12B | 12.2B | Runs well | FP16 | 24 GB | ~24 tok/s | |
| Gemma 3 12B | 12.2B | Runs well | FP16 | 24 GB | ~24 tok/s | |
| Gemma 2 9B | 9.2B | Runs well | FP16 | 19 GB | ~32 tok/s | |
| Qwen3 8B | 8.2B | Runs well | FP16 | 17 GB | ~36 tok/s | |
| Llama 3.1 8B | 8.0B | Runs well | FP16 | 16 GB | ~37 tok/s | |
| Qwen2.5-Coder 7B | 7.6B | Runs well | FP16 | 15 GB | ~39 tok/s | |
| Mistral 7B v0.3 | 7.3B | Runs well | FP16 | 15 GB | ~41 tok/s | |
| Gemma 3 4B | 4.3B | Runs well | FP16 | 9.1 GB | ~69 tok/s | |
| Qwen3 4B | 4B | Runs well | FP16 | 8.5 GB | ~74 tok/s | |
| Llama 3.2 3B | 3.2B | Runs well | FP16 | 7.0 GB | ~92 tok/s | |
| Llama 3.2 1B | 1.2B | Runs well | FP16 | 3.2 GB | ~238 tok/s | |
| DeepSeek-R1 671B-A37B (MoE)MoE | 671B | Runs (tight) | Q3_K_M | 318 GB | ~21 tok/s |
Estimates, computed in your browser — VRAM, quantisation and speed vary with your runtime and settings. How we estimate →
Top picks for the Mac Studio · M3 Ultra, 512 GB
Every model on the Mac Studio · M3 Ultra, 512 GB
| Model | Size | Fit | Best quant | Needs | Speed |
|---|---|---|---|---|---|
| Qwen3 235B-A22B (MoE)MoE | 235B | Runs well | Q8 | 244 GB | ~16 tok/s |
| Llama 4 Scout (MoE)MoE | 109B | Runs well | FP16 | 213 GB | ~11 tok/s |
| Qwen2.5 72B | 72.7B | Runs well | FP16 | 144 GB | ~4.1 tok/s |
| Llama 3.3 70B | 70.6B | Runs well | FP16 | 140 GB | ~4.2 tok/s |
| DeepSeek-R1 Distill Llama 70B | 70.6B | Runs well | FP16 | 140 GB | ~4.2 tok/s |
| Mixtral 8x7B (MoE)MoE | 46.7B | Runs well | FP16 | 92 GB | ~15 tok/s |
| Qwen3 32B | 32.8B | Runs well | FP16 | 66 GB | ~9.0 tok/s |
| Qwen2.5 32B | 32.8B | Runs well | FP16 | 66 GB | ~9.0 tok/s |
| Qwen2.5-Coder 32B | 32.8B | Runs well | FP16 | 66 GB | ~9.0 tok/s |
| DeepSeek-R1 Distill Qwen 32B | 32.8B | Runs well | FP16 | 66 GB | ~9.0 tok/s |
| Qwen3 30B-A3B (MoE)MoE | 30.5B | Runs well | FP16 | 61 GB | ~58 tok/s |
| Gemma 3 27B | 27.4B | Runs well | FP16 | 58 GB | ~11 tok/s |
| Gemma 2 27B | 27.2B | Runs well | FP16 | 56 GB | ~11 tok/s |
| Mistral Small 3 24B | 23.6B | Runs well | FP16 | 48 GB | ~12 tok/s |
| Qwen3 14B | 14.8B | Runs well | FP16 | 31 GB | ~20 tok/s |
| Qwen2.5 14B | 14.8B | Runs well | FP16 | 31 GB | ~20 tok/s |
| DeepSeek-R1 Distill Qwen 14B | 14.8B | Runs well | FP16 | 31 GB | ~20 tok/s |
| Phi-4 14B | 14.7B | Runs well | FP16 | 31 GB | ~20 tok/s |
| Mistral Nemo 12B | 12.2B | Runs well | FP16 | 26 GB | ~24 tok/s |
| Gemma 3 12B | 12.2B | Runs well | FP16 | 27 GB | ~24 tok/s |
| Gemma 2 9B | 9.2B | Runs well | FP16 | 21 GB | ~32 tok/s |
| Qwen3 8B | 8.2B | Runs well | FP16 | 18 GB | ~36 tok/s |
| Llama 3.1 8B | 8.0B | Runs well | FP16 | 17 GB | ~37 tok/s |
| Qwen2.5-Coder 7B | 7.6B | Runs well | FP16 | 16 GB | ~39 tok/s |
| Mistral 7B v0.3 | 7.3B | Runs well | FP16 | 16 GB | ~41 tok/s |
| Gemma 3 4B | 4.3B | Runs well | FP16 | 10 GB | ~69 tok/s |
| Qwen3 4B | 4B | Runs well | FP16 | 9.6 GB | ~74 tok/s |
| Llama 3.2 3B | 3.2B | Runs well | FP16 | 7.8 GB | ~92 tok/s |
| Llama 3.2 1B | 1.2B | Runs well | FP16 | 3.4 GB | ~238 tok/s |
| DeepSeek-R1 671B-A37B (MoE)MoE | 671B | Runs (tight) | Q3 | 349 GB | ~21 tok/s |
Similar hardware
FAQ
What is the best LLM for the Mac Studio · M3 Ultra, 512 GB?
For general use, Qwen3 235B-A22B (MoE) is the strongest model that runs well on the Mac Studio · M3 Ultra, 512 GB. See the picks-by-use-case below for coding, reasoning and more.
How much can the Mac Studio · M3 Ultra, 512 GB run?
The Mac Studio · M3 Ultra, 512 GB has 512 GB of unified memory, of which about 369 GB is usable for a model. That runs 29 of our tracked models well and 1 more at a tight quantisation.
Estimates at 8K context — see how we compute these.