Skip to content
mlx.app

M3 Ultra · 96 GB

819 GB/s memory bandwidth · 76.8 GB usable for models · 19.2 GB left to macOS

The short answer

36 models run comfortably and 1 more will run tight at 8K context. The largest that fits is Llama 3.2 90B Vision Instruct at q4, generating around 13.7 tokens per second.

macOS19.2 GBweights49.5 GBKV cache0.0 GBheadroom27.3 GB

Every model on this Mac

ModelSizeBest quantVerdictSpeed est.
Qwen2.5 0.5B Instruct490Mbf16FITS · 75.7 GB FREE694 tok/s
BGE-M3567Mbf16FITS · 75.7 GB FREE599 tok/s
Qwen3 0.6B600Mbf16FITS · 74.7 GB FREE566 tok/s
Parakeet TDT 0.6B v2600Mbf16FITS · 75.6 GB FREE566 tok/s
Llama 3.2 1B Instruct1.2Bbf16FITS · 74.1 GB FREE274 tok/s
Whisper Large v31.6Bbf16FITS · 73.7 GB FREE219 tok/s
Qwen3 1.7B1.7Bbf16FITS · 72.5 GB FREE200 tok/s
Llama 3.2 3B Instruct3.2Bbf16FITS · 69.9 GB FREE106 tok/s
Qwen3 4B4Bq8FITS · 71.3 GB FREE160 tok/s
Qwen3 Embedding 4B4Bq8FITS · 71.3 GB FREE160 tok/s
Mistral 7B Instruct v0.37.3Bq8FITS · 68.0 GB FREE87.6 tok/s
Qwen2.5 7B Instruct7.6Bbf16FITS · 61.1 GB FREE44.7 tok/s
InternVL3 8B8Bq8FITS · 68.3 GB FREE80.0 tok/s
Llama 3.1 8B Instruct8.0Bbf16FITS · 59.7 GB FREE42.3 tok/s
Qwen3 8B8.2Bq8FITS · 66.9 GB FREE78.0 tok/s
Qwen2.5-VL 7B Instruct8.3Bq8FITS · 67.5 GB FREE77.1 tok/s
Gemma 2 9B IT9.2Bq8FITS · 67.0 GB FREE69.2 tok/s
Llama 3.2 11B Vision Instruct10.7Bbf16FITS · 55.4 GB FREE31.8 tok/s
Qwen2.5 14B Instruct14.7Bq8FITS · 59.6 GB FREE43.5 tok/s
Qwen3 14B14.8Bq8FITS · 59.7 GB FREE43.2 tok/s
GPT-OSS 20B21Bq8FITS · 54.1 GB FREE178 tok/s
Codestral 22B v0.122.2Bq8FITS · 51.3 GB FREE28.8 tok/s
Mistral Small 24B Instruct 250124Bq8FITS · 50.0 GB FREE26.7 tok/s
Gemma 3 27B IT27Bbf16FITS · 22.8 GB FREE12.6 tok/s
Gemma 2 27B IT27.2Bq8FITS · 47.9 GB FREE23.5 tok/s
Qwen3 30B-A3B Instruct 250730.5Bq8FITS · 43.6 GB FREE194 tok/s
Qwen3 Coder 30B-A3B Instruct30.5Bq8FITS · 43.6 GB FREE194 tok/s
Qwen2.5 32B Instruct32.5Bq8FITS · 40.1 GB FREE19.7 tok/s
Qwen2.5 Coder 32B Instruct32.5Bq8FITS · 40.1 GB FREE19.7 tok/s
DeepSeek R1 Distill Qwen 32B32.5Bq8FITS · 40.1 GB FREE19.7 tok/s
Qwen3 32B32.8Bq8FITS · 39.8 GB FREE19.5 tok/s
Qwen2.5-VL 32B Instruct33Bq8FITS · 39.6 GB FREE19.4 tok/s
Mixtral 8x7B Instruct v0.146.7Bq8FITS · 26.1 GB FREE49.6 tok/s
Llama 3.3 70B Instruct70.6Bq4FITS · 34.4 GB FREE17.1 tok/s
Qwen2.5 72B Instruct72.7Bq4FITS · 33.2 GB FREE16.6 tok/s
Llama 3.2 90B Vision Instruct88Bq4FITS · 27.3 GB FREE13.7 tok/s
GPT-OSS 120B116.8Bq4TIGHT · 10.5 GB FREE237 tok/s
Qwen3 235B-A22B235BOVER
DeepSeek V3671BOVER