Log inCreate wallet

What can an M2 Ultra with 128 GB run?

47 of 73 open models run well on it. The GPU can use about 96 GB of its 128 GB by default, and it reads memory at 800 GB/s, which sets how fast it writes. Found in: Mac Studio (2023), Mac Pro (2023), Mac Pro (Rack, 2023).

Smartest that runs wellQwen3.8 27B8-bit · 28.7 GB · 13–22 tok/sBiggest that runs wellQwen3.5 122B A10B4-bit · 68.6 GB · 64–114 tok/sFor codeQwen3 Coder Next8-bit · 85 GB · 113–200 tok/s

Runs well 51

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

ModelVersionMemoryTokens/s
Qwen3.8 27B27B · Intelligence 33.78-bit28.7 GB13–22
Nemotron 3.5 Bolt 74B A9B74B · 9B active · Intelligence 21.48-bit78.6 GB38–67
Qwen3.6 35B A3B35B · 3B active · Intelligence 18.28-bit37.2 GB113–200
Muse Glimmer 30B29.6B · Intelligence 17.58-bit31.5 GB11–20
Qwen3.5 122B A10B122B · 10B active · Intelligence 15.64-bit68.6 GB64–114
Gemma 4 31B30.7B · Intelligence 14.78-bit32.6 GB11–20
Nemotron 3.5 Lightning 30B A3B30B · 3B active · Intelligence 12.98-bit31.9 GB113–200
Granite 4.2 30B30B · Intelligence 12.78-bit31.9 GB11–20
MiniCPM5 2B2.52B · Intelligence 12.58-bit2.7 GB133–237
gpt-oss-120b117B · 5.1B active · Intelligence 11.6native 4.25-bit65.3 GB133–237
Mistral Small 4 (119B)119B · 6.5B active · Intelligence 11.34-bit66.9 GB97–173
Qwen3.5 9B9B · Intelligence 11.28-bit9.6 GB38–67
Granite 4.2 8B8B · Intelligence 11.18-bit8.5 GB42–75
Qwen3 Coder Next80B · 3B active · Intelligence 9.28-bit85 GB113–200
Granite 4.2 3B3B · Intelligence 9.18-bit3.2 GB113–200
gpt-oss-20b21B · 3.6B active · Intelligence 9native 4.25-bit13.8 GB189–337
Nemotron 3 Nano 30B A3B30B · 3.5B active · Intelligence 8.98-bit31.9 GB97–173
Gemma 4 E4B8B · Intelligence 8.98-bit8.5 GB75–133
Llama 4 Scout 17B-16E Instruct109B · 17B active · Intelligence 6.56-bit88.6 GB26–46
Ministral 3 8B8.9B · Intelligence 5.58-bit9.5 GB38–67
Nemotron 3 Nano 4B3.97B · Intelligence 4.88-bit4.2 GB86–152
Ornith 1.5 122B122B · 10B active4-bit68.6 GB64–114
Nemotron 3 Super 120B A12B120B · 12B active4-bit67.5 GB53–94
GLM-4.5 Air106B · 12B active6-bit86.1 GB37–65
Qwen3 Coder 30B A3B Instruct30.5B · 3.3B active8-bit32.4 GB103–183
GLM-4.7 Flash30B · 3B active8-bit31.9 GB113–200
Qwen3.6 27B27B8-bit28.7 GB13–22
Gemma 4 26B A4B25.2B · 3.8B active8-bit26.8 GB90–160
Devstral Small 2 (24B)24B8-bit25.5 GB14–25
Magistral Small 1.2 (2509)24B8-bit25.5 GB14–25
Mistral Small 3.2 (24B)24B8-bit25.5 GB14–25
LFM2.5 24B A2B24B · 2B active8-bit25.5 GB171–305
Phi-4 Reasoning Vision 15B15B8-bit15.9 GB23–40
Phi-4 (14B)14B8-bit14.9 GB24–43
Gemma 4 12B Unified12B8-bit12.7 GB28–50
Qwen3 8B8.2B8-bit8.7 GB41–74
Llama 3.1 8B Instruct8B8-bit8.5 GB42–75
Gemma 4 E2B5.1B8-bit5.4 GB150–267
Qwen3.5 4B4B8-bit4.3 GB84–149
Phi-4 mini instruct3.8B8-bit4 GB90–160
Llama 3.2 3B Instruct3.21B8-bit3.4 GB106–188
SmolLM3 3B3B8-bit3.2 GB113–200
LFM2.5 2.6B2.69B8-bit2.9 GB124–221
Qwen3.5 2B2B8-bit2.1 GB171–305
Llama 3.2 1B Instruct1.23B8-bit1.3 GB277–492
Qwen3.5 0.8B0.8B8-bit0.9 GB400–711
Qwen3 0.6B0.6B8-bit0.6 GB600–1067
Qwen3.8 Flash Next180B · 6B active · Intelligence 39.83-bit lower quality78.8 GB138–246
Inkling Small276B · 12B active · Intelligence 25.72-bit lower quality86.3 GB95–168
DeepSeek V4 Flash 0731284B · 13B active2-bit lower quality88.8 GB88–156
MiniMax M2.7230B · 10B active2-bit lower quality71.9 GB116–206

Runs, but slowly 3

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

ModelVersionMemoryTokens/s
Mistral Medium 3.5 (128B)128B · Intelligence 14.24-bit72 GB5–9
Devstral 2 (123B)123B4-bit69.2 GB5–9
Llama 3.3 70B Instruct70B8-bit74.4 GB5–9

Only just fits 3

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

ModelVersionMemoryTokens/s
GLM-5.3 Flash320B · 18B active · Intelligence 41.82-bit lower quality100 GB64–114
MiMo V2.6 Flash309B · 15B active · Intelligence 37.92-bit lower quality96.6 GB77–136
Hy3295B · 21B active · Intelligence 25.32-bit lower quality92.2 GB55–97

Too big 16

Needs more memory than 128 GB, even at 2 bits.

Other Macs

Every Mac with 128 GB, from any chip.

The M2 Ultra with more or less memory:

Neighbouring chips:

Head to head:

Compare it with any Mac or GPU · The M2 Ultra for local AI · The best Mac for local LLMs · Every Mac

Questions

What is the best LLM for an M2 Ultra with 128 GB?
Of the models with a published Intelligence Index, Qwen3.8 27B scores highest among those that run well here: 8-bit, about 13–22 tokens a second. The biggest that runs well is Qwen3.5 122B A10B.
Can an M2 Ultra with 128 GB run Llama 3.3 70B?
Yes, but slowly: 8-bit, at about 5–9 tokens a second.
How fast is gpt-oss-20b on an M2 Ultra?
About 189–337 tokens a second (native 4.25-bit), from the chip’s 800 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 M2 Ultra with 128 GB run? Models and speed