With 96 GB unified memory (about 72.0 GB usable by the GPU), these are the best open models that fit. The top pick is Llama 4 Scout 109B (MoE) at Q4_K_M (~69.3 GB, ~31 tok/s): ollama run llama4:16x17b.
Best for coding: gpt-oss 120B (MoE, MXFP4) (gpt-oss:120b), Qwen3-Coder 30B-A3B (MoE) (qwen3-coder:30b-a3b-q8_0), Qwen3 30B-A3B (MoE) (qwen3:30b-a3b-q8_0).
Runtimes available for this hardware: Ollama, LM Studio, llama.cpp, MLX.
Open the interactive checker for Mac Studio M3 Ultra 96GB| # | Model | Quant | Memory needed | Fit | Speed | Command |
|---|---|---|---|---|---|---|
| 1 | Llama 4 Scout 109B (MoE) | Q4_K_M | 69.3 GB | fits in fast memory | ~31 tok/s | ollama run llama4:16x17b |
| 2 | gpt-oss 120B (MoE, MXFP4) | MXFP4 | 67.3 GB | fits in fast memory | ~116 tok/s | ollama run gpt-oss:120b |
| 3 | Qwen3-Next 80B-A3B (MoE) | Q4_K_M | 52.4 GB | fits in fast memory | ~175 tok/s | ollama run qwen3-next:80b |
| 4 | Llama 3.3 70B | Q5_K_M | 54.1 GB | fits in fast memory | ~13 tok/s | ollama run llama3.3:70b-instruct-q5_K_M |
| 5 | DeepSeek R1 Distill Llama 70B | Q4_K_M | 47.1 GB | fits in fast memory | ~15 tok/s | ollama run deepseek-r1:70b |
| 6 | Qwen3 30B-A3B (MoE) | Q8_0 | 34.4 GB | fits in fast memory | ~92 tok/s | ollama run qwen3:30b-a3b-q8_0 |
| 7 | Qwen3-Coder 30B-A3B (MoE) | Q8_0 | 33.4 GB | fits in fast memory | ~95 tok/s | ollama run qwen3-coder:30b-a3b-q8_0 |
| 8 | gpt-oss 20B (MoE, MXFP4) | MXFP4 | 15.1 GB | fits in fast memory | ~136 tok/s | ollama run gpt-oss:20b |
| 9 | Qwen3 32B | Q8_0 | 37.2 GB | fits in fast memory | ~19 tok/s | ollama run qwen3:32b-q8_0 |
| 10 | DeepSeek R1 Distill Qwen 32B | Q8_0 | 37.2 GB | fits in fast memory | ~19 tok/s | ollama run deepseek-r1:32b-qwen-distill-q8_0 |
| 11 | Gemma 3 27B | Q8_0 | 32 GB | fits in fast memory | ~22 tok/s | ollama run gemma3:27b-it-q8_0 |
| 12 | Clef 27B (Cloudflare, decision) | Q8_0 | 31.9 GB | fits in fast memory | ~22 tok/s | ollama run clef:27b-q8_0 |
Estimates, not benchmarks. Model sizes are ollama.com download sizes; memory needed adds headroom for ~8k tokens of context; speeds are rough figures from memory bandwidth. Data updated 2026-10-14.
A project by OnDevice Labs · llms.txt