Log inCreate wallet

What can an M5 Max with 36 GB run?

38 of 73 open models run well on it. The GPU can use about 27 GB of its 36 GB by default, and it reads memory at 460 GB/s, which sets how fast it writes. Found in: MacBook Pro (14-inch, M5 Pro or M5 Max), MacBook Pro (16-inch, M5 Pro or M5 Max), Mac Studio (M5 Max or M5 Ultra).

Smartest that runs wellQwen3.8 27B4-bit · 15.2 GB · 14–24 tok/sBiggest that runs wellQwen3.6 35B A3B4-bit · 19.7 GB · 122–216 tok/sFor codeQwen3 Coder 30B A3B Instruct6-bit · 24.8 GB · 77–136 tok/s

Runs well 40

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

ModelVersionMemoryTokens/s
Qwen3.8 27B27B · Intelligence 33.74-bit15.2 GB14–24
Qwen3.6 35B A3B35B · 3B active · Intelligence 18.24-bit19.7 GB122–216
Muse Glimmer 30B29.6B · Intelligence 17.54-bit16.7 GB12–22
Gemma 4 31B30.7B · Intelligence 14.74-bit17.3 GB12–21
Nemotron 3.5 Lightning 30B A3B30B · 3B active · Intelligence 12.96-bit24.4 GB86–153
Granite 4.2 30B30B · Intelligence 12.74-bit16.9 GB12–22
MiniCPM5 2B2.52B · Intelligence 12.58-bit2.7 GB77–136
Qwen3.5 9B9B · Intelligence 11.28-bit9.6 GB22–38
Granite 4.2 8B8B · Intelligence 11.18-bit8.5 GB24–43
Granite 4.2 3B3B · Intelligence 9.18-bit3.2 GB65–115
gpt-oss-20b21B · 3.6B active · Intelligence 9native 4.25-bit13.8 GB109–194
Nemotron 3 Nano 30B A3B30B · 3.5B active · Intelligence 8.96-bit24.4 GB74–131
Gemma 4 E4B8B · Intelligence 8.98-bit8.5 GB43–77
Ministral 3 8B8.9B · Intelligence 5.58-bit9.5 GB22–39
Nemotron 3 Nano 4B3.97B · Intelligence 4.88-bit4.2 GB49–88
Qwen3 Coder 30B A3B Instruct30.5B · 3.3B active6-bit24.8 GB77–136
GLM-4.7 Flash30B · 3B active6-bit24.4 GB86–153
Qwen3.6 27B27B4-bit15.2 GB14–24
Gemma 4 26B A4B25.2B · 3.8B active6-bit20.5 GB67–119
Devstral Small 2 (24B)24B6-bit19.5 GB11–19
Magistral Small 1.2 (2509)24B6-bit19.5 GB11–19
Mistral Small 3.2 (24B)24B6-bit19.5 GB11–19
LFM2.5 24B A2B24B · 2B active6-bit19.5 GB129–230
Phi-4 Reasoning Vision 15B15B8-bit15.9 GB13–23
Phi-4 (14B)14B8-bit14.9 GB14–25
Gemma 4 12B Unified12B8-bit12.7 GB16–29
Qwen3 8B8.2B8-bit8.7 GB24–42
Llama 3.1 8B Instruct8B8-bit8.5 GB24–43
Gemma 4 E2B5.1B8-bit5.4 GB86–153
Qwen3.5 4B4B8-bit4.3 GB48–86
Phi-4 mini instruct3.8B8-bit4 GB52–92
Llama 3.2 3B Instruct3.21B8-bit3.4 GB61–108
SmolLM3 3B3B8-bit3.2 GB65–115
LFM2.5 2.6B2.69B8-bit2.9 GB71–127
Qwen3.5 2B2B8-bit2.1 GB99–175
Llama 3.2 1B Instruct1.23B8-bit1.3 GB159–283
Qwen3.5 0.8B0.8B8-bit0.9 GB230–409
Qwen3 0.6B0.6B8-bit0.6 GB345–613
Nemotron 3.5 Bolt 74B A9B74B · 9B active · Intelligence 21.42-bit lower quality23.1 GB74–131
Qwen3 Coder Next80B · 3B active · Intelligence 9.22-bit lower quality25 GB230–409

Runs, but slowly 1

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

ModelVersionMemoryTokens/s
Llama 3.3 70B Instruct70B2-bit lower quality21.9 GB9–17

Too big 32

Needs more memory than 36 GB, even at 2 bits.

Other Macs

Every Mac with 36 GB, from any chip.

The M5 Max with more or less memory:

Neighbouring chips:

Head to head:

Against PCs and graphics cards:

Compare it with any Mac or GPU · The M5 Max for local AI · The best Mac for local LLMs · Every Mac

Questions

What is the best LLM for an M5 Max with 36 GB?
Of the models with a published Intelligence Index, Qwen3.8 27B scores highest among those that run well here: 4-bit, about 14–24 tokens a second. The biggest that runs well is Qwen3.6 35B A3B.
Can an M5 Max with 36 GB run Llama 3.3 70B?
Yes, but slowly: 2-bit (lower quality), at about 9–17 tokens a second.
How fast is gpt-oss-20b on an M5 Max?
About 109–194 tokens a second (native 4.25-bit), from the chip’s 460 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 M5 Max with 36 GB run? Models and speed