Log inCreate wallet

What can an M4 Max with 48 GB run?

38 of 73 open models run well on it. The GPU can use about 36 GB of its 48 GB by default, and it reads memory at 546 GB/s, which sets how fast it writes. Found in: MacBook Pro (14-inch, M4 Pro or M4 Max, 2024), MacBook Pro (16-inch, 2024), Mac Studio (2025).

Smartest that runs wellQwen3.8 27B6-bit · 21.9 GB · 11–20 tok/sBiggest that runs wellQwen3.6 35B A3B6-bit · 28.4 GB · 102–182 tok/sFor codeQwen3 Coder 30B A3B Instruct8-bit · 32.4 GB · 70–125 tok/s

Runs well 41

Fits with room for context, at 10 or more tokens a second.

ModelVersionMemoryTokens/s
Qwen3.8 27B27B · Intelligence 33.76-bit21.9 GB11–20
Qwen3.6 35B A3B35B · 3B active · Intelligence 18.26-bit28.4 GB102–182
Muse Glimmer 30B29.6B · Intelligence 17.56-bit24.1 GB10–18
Gemma 4 31B30.7B · Intelligence 14.76-bit24.9 GB10–18
Nemotron 3.5 Lightning 30B A3B30B · 3B active · Intelligence 12.98-bit31.9 GB77–137
Granite 4.2 30B30B · Intelligence 12.76-bit24.4 GB10–18
MiniCPM5 2B2.52B · Intelligence 12.58-bit2.7 GB91–162
Qwen3.5 9B9B · Intelligence 11.28-bit9.6 GB26–46
Granite 4.2 8B8B · Intelligence 11.18-bit8.5 GB29–51
Granite 4.2 3B3B · Intelligence 9.18-bit3.2 GB77–137
gpt-oss-20b21B · 3.6B active · Intelligence 9native 4.25-bit13.8 GB129–230
Nemotron 3 Nano 30B A3B30B · 3.5B active · Intelligence 8.98-bit31.9 GB66–118
Gemma 4 E4B8B · Intelligence 8.98-bit8.5 GB51–91
Ministral 3 8B8.9B · Intelligence 5.58-bit9.5 GB26–46
Nemotron 3 Nano 4B3.97B · Intelligence 4.88-bit4.2 GB59–104
Qwen3 Coder 30B A3B Instruct30.5B · 3.3B active8-bit32.4 GB70–125
GLM-4.7 Flash30B · 3B active8-bit31.9 GB77–137
Qwen3.6 27B27B6-bit21.9 GB11–20
Gemma 4 26B A4B25.2B · 3.8B active8-bit26.8 GB61–109
Devstral Small 2 (24B)24B8-bit25.5 GB10–17
Magistral Small 1.2 (2509)24B8-bit25.5 GB10–17
Mistral Small 3.2 (24B)24B8-bit25.5 GB10–17
LFM2.5 24B A2B24B · 2B active8-bit25.5 GB117–208
Phi-4 Reasoning Vision 15B15B8-bit15.9 GB15–27
Phi-4 (14B)14B8-bit14.9 GB16–29
Gemma 4 12B Unified12B8-bit12.7 GB19–34
Qwen3 8B8.2B8-bit8.7 GB28–50
Llama 3.1 8B Instruct8B8-bit8.5 GB29–51
Gemma 4 E2B5.1B8-bit5.4 GB102–182
Qwen3.5 4B4B8-bit4.3 GB57–102
Phi-4 mini instruct3.8B8-bit4 GB61–109
Llama 3.2 3B Instruct3.21B8-bit3.4 GB72–128
SmolLM3 3B3B8-bit3.2 GB77–137
LFM2.5 2.6B2.69B8-bit2.9 GB85–151
Qwen3.5 2B2B8-bit2.1 GB117–208
Llama 3.2 1B Instruct1.23B8-bit1.3 GB189–336
Qwen3.5 0.8B0.8B8-bit0.9 GB273–485
Qwen3 0.6B0.6B8-bit0.6 GB410–728
Nemotron 3.5 Bolt 74B A9B74B · 9B active · Intelligence 21.43-bit lower quality32.4 GB63–112
Qwen3 Coder Next80B · 3B active · Intelligence 9.22-bit lower quality25 GB273–485
GLM-4.5 Air106B · 12B active2-bit lower quality33.1 GB65–115

Runs, but slowly 1

Fits, but under 10 tokens a second: fine for batch work, slow for chat.

ModelVersionMemoryTokens/s
Llama 3.3 70B Instruct70B3-bit lower quality30.6 GB8–14

Only just fits 8

Needs almost all memory: raise the GPU memory limit and keep the context short.

ModelVersionMemoryTokens/s
Qwen3.5 122B A10B122B · 10B active · Intelligence 15.62-bit lower quality38.1 GB79–141
Mistral Medium 3.5 (128B)128B · Intelligence 14.22-bit lower quality40 GB6–11
gpt-oss-120b117B · 5.1B active · Intelligence 11.62-bit lower quality36.6 GB154–273
Mistral Small 4 (119B)119B · 6.5B active · Intelligence 11.32-bit lower quality37.2 GB123–218
Llama 4 Scout 17B-16E Instruct109B · 17B active · Intelligence 6.52-bit lower quality34.1 GB46–82
Devstral 2 (123B)123B2-bit lower quality38.4 GB6–11
Ornith 1.5 122B122B · 10B active2-bit lower quality38.1 GB79–141
Nemotron 3 Super 120B A12B120B · 12B active2-bit lower quality37.5 GB65–115

Too big 23

Needs more memory than 48 GB, even at 2 bits.

Other Macs

Every Mac with 48 GB, from any chip.

The M4 Max with more or less memory:

Neighbouring chips:

Head to head:

Against PCs and graphics cards:

Compare it with any Mac or GPU · The M4 Max for local AI · The best Mac for local LLMs · Every Mac

Questions

What is the best LLM for an M4 Max with 48 GB?
Of the models with a published Intelligence Index, Qwen3.8 27B scores highest among those that run well here: 6-bit, about 11–20 tokens a second. The biggest that runs well is Qwen3.6 35B A3B.
Can an M4 Max with 48 GB run Llama 3.3 70B?
Yes, but slowly: 3-bit (lower quality), at about 8–14 tokens a second.
How fast is gpt-oss-20b on an M4 Max?
About 129–230 tokens a second (native 4.25-bit), from the chip’s 546 GB/s memory bandwidth.

Run them with MLX Harness (free, open source) or Apple’s mlx-lm in Terminal: step by step.

How it’s worked out

Memory: the model’s parameters at 8, 6, 4, 3 or 2 bits per weight plus MLX’s group scales (half a bit per weight), or its native format, with room for the context on top. By default macOS lets the GPU use about 75% of unified memory. Speed: memory bandwidth divided by the bytes read per token is the ceiling for one conversation; in llama.cpp’s public Apple Silicon benchmarks real runs reach about 45% to 80% of it, which is the range shown. These are estimates, not measurements: measure your own Mac on MLX Bench. Model figures from each official model card, chip figures from Apple’s Tech Specs. Checked 2026-10-19. Same data as mlx.app/compare.

What LLMs can an M4 Max with 48 GB run? Models and speed