With 48 GB unified memory (about 36.0 GB usable by the GPU), these are the best open models that fit. The top pick is Qwen3 30B-A3B (MoE) at Q8_0 (~34.4 GB, ~69 tok/s): ollama run qwen3:30b-a3b-q8_0.
Best for coding: Qwen3-Coder 30B-A3B (MoE) (qwen3-coder:30b-a3b-q8_0), Qwen3 30B-A3B (MoE) (qwen3:30b-a3b-q8_0), gpt-oss 20B (MoE, MXFP4) (gpt-oss:20b).
Runtimes available for this hardware: Ollama, LM Studio, llama.cpp, MLX.
Open the interactive checker for Mac Studio M5 Max (40c GPU) 48GB| # | Model | Quant | Memory needed | Fit | Speed | Command |
|---|---|---|---|---|---|---|
| 1 | Qwen3 30B-A3B (MoE) | Q8_0 | 34.4 GB | fits in fast memory | ~69 tok/s | ollama run qwen3:30b-a3b-q8_0 |
| 2 | Qwen3-Coder 30B-A3B (MoE) | Q8_0 | 33.4 GB | fits in fast memory | ~71 tok/s | ollama run qwen3-coder:30b-a3b-q8_0 |
| 3 | gpt-oss 20B (MoE, MXFP4) | MXFP4 | 15.1 GB | fits in fast memory | ~102 tok/s | ollama run gpt-oss:20b |
| 4 | Qwen3 32B | Q4_K_M | 22.2 GB | fits in fast memory | ~25 tok/s | ollama run qwen3:32b |
| 5 | DeepSeek R1 Distill Qwen 32B | Q4_K_M | 22.2 GB | fits in fast memory | ~25 tok/s | ollama run deepseek-r1:32b |
| 6 | Gemma 3 27B | Q8_0 | 32 GB | fits in fast memory | ~16 tok/s | ollama run gemma3:27b-it-q8_0 |
| 7 | Clef 27B (Cloudflare, decision) | Q8_0 | 31.9 GB | fits in fast memory | ~16 tok/s | ollama run clef:27b-q8_0 |
| 8 | Qwen3.8 27B | Q8_0 | 31.9 GB | fits in fast memory | ~16 tok/s | ollama run qwen3.8:27b-q8_0 |
| 9 | Mistral Small 24B | Q8_0 | 26.8 GB | fits in fast memory | ~20 tok/s | ollama run mistral-small:24b-instruct-2501-q8_0 |
| 10 | Mixtral 8x7B (MoE) | Q5_K_M | 34.8 GB | fits in fast memory | ~27 tok/s | ollama run mixtral:8x7b-instruct-v0.1-q5_K_M |
| 11 | Qwen3 14B | Q8_0 | 17.3 GB | fits in fast memory | ~31 tok/s | ollama run qwen3:14b-q8_0 |
| 12 | DeepSeek R1 Distill Qwen 14B | Q8_0 | 17.3 GB | fits in fast memory | ~31 tok/s | ollama run deepseek-r1:14b-qwen-distill-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