Log inCreate wallet

What can an M3 Ultra with 96 GB run?

43 of 73 open models run well on it. The GPU can use about 72 GB of its 96 GB by default, and it reads memory at 819 GB/s, which sets how fast it writes. Found in: Mac Studio (2025).

Smartest that runs wellQwen3.8 27B8-bit · 28.7 GB · 13–23 tok/sBiggest that runs wellgpt-oss-120bnative 4.25-bit · 65.3 GB · 137–243 tok/sFor codeQwen3 Coder Next6-bit · 65 GB · 154–273 tok/s

Runs well 48

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

ModelVersionMemoryTokens/s
Qwen3.8 27B27B · Intelligence 33.78-bit28.7 GB13–23
Nemotron 3.5 Bolt 74B A9B74B · 9B active · Intelligence 21.46-bit60.1 GB50–90
Qwen3.6 35B A3B35B · 3B active · Intelligence 18.28-bit37.2 GB115–205
Muse Glimmer 30B29.6B · Intelligence 17.58-bit31.5 GB12–21
Gemma 4 31B30.7B · Intelligence 14.78-bit32.6 GB11–20
Nemotron 3.5 Lightning 30B A3B30B · 3B active · Intelligence 12.98-bit31.9 GB115–205
Granite 4.2 30B30B · Intelligence 12.78-bit31.9 GB12–21
MiniCPM5 2B2.52B · Intelligence 12.58-bit2.7 GB137–243
gpt-oss-120b117B · 5.1B active · Intelligence 11.6native 4.25-bit65.3 GB137–243
Qwen3.5 9B9B · Intelligence 11.28-bit9.6 GB38–68
Granite 4.2 8B8B · Intelligence 11.18-bit8.5 GB43–77
Qwen3 Coder Next80B · 3B active · Intelligence 9.26-bit65 GB154–273
Granite 4.2 3B3B · Intelligence 9.18-bit3.2 GB115–205
gpt-oss-20b21B · 3.6B active · Intelligence 9native 4.25-bit13.8 GB194–345
Nemotron 3 Nano 30B A3B30B · 3.5B active · Intelligence 8.98-bit31.9 GB100–177
Gemma 4 E4B8B · Intelligence 8.98-bit8.5 GB77–137
Llama 4 Scout 17B-16E Instruct109B · 17B active · Intelligence 6.54-bit61.3 GB38–68
Ministral 3 8B8.9B · Intelligence 5.58-bit9.5 GB39–69
Nemotron 3 Nano 4B3.97B · Intelligence 4.88-bit4.2 GB88–156
GLM-4.5 Air106B · 12B active4-bit59.6 GB54–96
Qwen3 Coder 30B A3B Instruct30.5B · 3.3B active8-bit32.4 GB105–187
GLM-4.7 Flash30B · 3B active8-bit31.9 GB115–205
Qwen3.6 27B27B8-bit28.7 GB13–23
Gemma 4 26B A4B25.2B · 3.8B active8-bit26.8 GB92–164
Devstral Small 2 (24B)24B8-bit25.5 GB14–26
Magistral Small 1.2 (2509)24B8-bit25.5 GB14–26
Mistral Small 3.2 (24B)24B8-bit25.5 GB14–26
LFM2.5 24B A2B24B · 2B active8-bit25.5 GB176–312
Phi-4 Reasoning Vision 15B15B8-bit15.9 GB23–41
Phi-4 (14B)14B8-bit14.9 GB25–44
Gemma 4 12B Unified12B8-bit12.7 GB29–52
Qwen3 8B8.2B8-bit8.7 GB42–75
Llama 3.1 8B Instruct8B8-bit8.5 GB43–77
Gemma 4 E2B5.1B8-bit5.4 GB154–273
Qwen3.5 4B4B8-bit4.3 GB86–152
Phi-4 mini instruct3.8B8-bit4 GB92–164
Llama 3.2 3B Instruct3.21B8-bit3.4 GB108–193
SmolLM3 3B3B8-bit3.2 GB115–205
LFM2.5 2.6B2.69B8-bit2.9 GB127–226
Qwen3.5 2B2B8-bit2.1 GB176–312
Llama 3.2 1B Instruct1.23B8-bit1.3 GB284–504
Qwen3.5 0.8B0.8B8-bit0.9 GB410–728
Qwen3 0.6B0.6B8-bit0.6 GB614–1092
Qwen3.8 Flash Next180B · 6B active · Intelligence 39.82-bit lower quality56.3 GB194–345
Qwen3.5 122B A10B122B · 10B active · Intelligence 15.63-bit lower quality53.4 GB84–149
Mistral Small 4 (119B)119B · 6.5B active · Intelligence 11.33-bit lower quality52.1 GB132–234
Ornith 1.5 122B122B · 10B active3-bit lower quality53.4 GB84–149
Nemotron 3 Super 120B A12B120B · 12B active3-bit lower quality52.5 GB70–124

Runs, but slowly 3

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

ModelVersionMemoryTokens/s
Llama 3.3 70B Instruct70B6-bit56.9 GB6–12
Mistral Medium 3.5 (128B)128B · Intelligence 14.23-bit lower quality56 GB7–12
Devstral 2 (123B)123B3-bit lower quality53.8 GB7–12

Only just fits 1

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

ModelVersionMemoryTokens/s
MiniMax M2.7230B · 10B active2-bit lower quality71.9 GB119–211

Too big 21

Needs more memory than 96 GB, even at 2 bits.

Other Macs

Every Mac with 96 GB, from any chip.

The M3 Ultra with more or less memory:

Neighbouring chips:

Head to head:

Against PCs and graphics cards:

Compare it with any Mac or GPU · The M3 Ultra for local AI · The best Mac for local LLMs · Every Mac

Questions

What is the best LLM for an M3 Ultra with 96 GB?
Of the models with a published Intelligence Index, Qwen3.8 27B scores highest among those that run well here: 8-bit, about 13–23 tokens a second. The biggest that runs well is gpt-oss-120b.
Can an M3 Ultra with 96 GB run Llama 3.3 70B?
Yes, but slowly: 6-bit, at about 6–12 tokens a second.
How fast is gpt-oss-20b on an M3 Ultra?
About 194–345 tokens a second (native 4.25-bit), from the chip’s 819 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 M3 Ultra with 96 GB run? Models and speed