FR
← All hardware
Unified memory · Apple unified memory

Mac Studio M3 Ultra 512 GB: which LLMs can it run?

Of 512 GB of unified memory, about 384 GB are usable by the GPU. On its own, it loads up to Llama 3.1 405B in Q4. It is a complete machine: you cannot add cards to it.

Specifications

Memory512 GB
Bandwidth819 GB/s
FP16 compute~57 TF
Power270 W
Indicative price€11,500
€ per usable GB€30
CompareSee prices ↗

Against cards at the same price

Mac Studio M3 Ultra 512 GB vs A100 80 GB€11,000
Mac Studio M3 Ultra 512 GB vs RTX PRO 6000 Blackwell 96 GB€9,000
Mac Studio M3 Ultra 512 GB vs L40S 48 GB€7,500

Recipes with this hardware

Qwen3 235B on a single Mac€11,500

Models and expected speed

ModelQ4Q8FP16
Qwen3 4B✓ ~177 tok/s✓ ~103 tok/s✓ ~55 tok/s
Llama 3.1 8B✓ ~91 tok/s✓ ~52 tok/s✓ ~28 tok/s
Qwen3 8B✓ ~89 tok/s✓ ~51 tok/s✓ ~27 tok/s
Gemma 3 12B✓ ~61 tok/s✓ ~35 tok/s✓ ~19 tok/s
Mistral Nemo 12B✓ ~60 tok/s✓ ~34 tok/s✓ ~18 tok/s
Phi-4 14B✓ ~50 tok/s✓ ~29 tok/s✓ ~15 tok/s
Qwen3 14B✓ ~50 tok/s✓ ~28 tok/s✓ ~15 tok/s
gpt-oss-20b✓ ~195 tok/s✓ ~114 tok/s✓ ~61 tok/s
Mistral Small 24B✓ ~31 tok/s✓ ~18 tok/s✓ ~9.4 tok/s
Gemma 3 27B✓ ~28 tok/s✓ ~16 tok/s✓ ~8.3 tok/s
Qwen3.8 27B✓ ~27 tok/s✓ ~15 tok/s✓ ~8.0 tok/s
Qwen3 30B-A3B✓ ~211 tok/s✓ ~123 tok/s✓ ~67 tok/s
Qwen3 32B✓ ~23 tok/s✓ ~13 tok/s✓ ~6.9 tok/s
Qwen2.5 Coder 32B✓ ~23 tok/s✓ ~13 tok/s✓ ~6.9 tok/s
DeepSeek R1 Distill 32B✓ ~23 tok/s✓ ~13 tok/s✓ ~6.9 tok/s
Mixtral 8x7B✓ ~57 tok/s✓ ~33 tok/s✓ ~17 tok/s
Llama 3.3 70B✓ ~11 tok/s✓ ~6.0 tok/s✓ ~3.2 tok/s
GLM-4.5-Air 106B✓ ~61 tok/s✓ ~35 tok/s✓ ~19 tok/s
gpt-oss-120b✓ ~140 tok/s✓ ~81 tok/s✓ ~44 tok/s
Mistral Large 123B✓ ~6.1 tok/s✓ ~3.5 tok/s✓ ~1.8 tok/s
Qwen3 235B-A22B✓ ~34 tok/s✓ ~19 tok/s✗
GLM-4.5 355B✓ ~23 tok/s✗✗
Llama 3.1 405B✓ ~1.9 tok/s✗✗
DeepSeek R1 671B✗✗✗
DeepSeek V3.1 671B✗✗✗
Kimi K2 1T✗✗✗

8k tokens, one user. “2 ×” = number of cards needed, speed with tensor parallelism (vLLM). ✗ = doesn’t fit, even with 1 cards.