Unified memory · Apple unified memory
Mac Studio M4 Max 128 GB: which LLMs can it run?
Of 128 GB of unified memory, about 96 GB are usable by the GPU. On its own, it loads up to Mistral Large 123B in Q4. It is a complete machine: you cannot add cards to it.
Specifications
Memory128 GB
Bandwidth546 GB/s
FP16 compute~34 TF
Power150 W
Indicative price€4,400
€ per usable GB€46
Against cards at the same price
Recipes with this hardware
Models and expected speed
| Model | Q4 | Q8 | FP16 |
|---|---|---|---|
| Qwen3 4B | ✓ ~118 tok/s | ✓ ~68 tok/s | ✓ ~37 tok/s |
| Llama 3.1 8B | ✓ ~61 tok/s | ✓ ~35 tok/s | ✓ ~19 tok/s |
| Qwen3 8B | ✓ ~59 tok/s | ✓ ~34 tok/s | ✓ ~18 tok/s |
| Gemma 3 12B | ✓ ~41 tok/s | ✓ ~23 tok/s | ✓ ~12 tok/s |
| Mistral Nemo 12B | ✓ ~40 tok/s | ✓ ~23 tok/s | ✓ ~12 tok/s |
| Phi-4 14B | ✓ ~33 tok/s | ✓ ~19 tok/s | ✓ ~10 tok/s |
| Qwen3 14B | ✓ ~33 tok/s | ✓ ~19 tok/s | ✓ ~10 tok/s |
| gpt-oss-20b | ✓ ~130 tok/s | ✓ ~76 tok/s | ✓ ~41 tok/s |
| Mistral Small 24B | ✓ ~21 tok/s | ✓ ~12 tok/s | ✓ ~6.2 tok/s |
| Gemma 3 27B | ✓ ~18 tok/s | ✓ ~10 tok/s | ✓ ~5.5 tok/s |
| Qwen3.8 27B | ✓ ~18 tok/s | ✓ ~10 tok/s | ✓ ~5.3 tok/s |
| Qwen3 30B-A3B | ✓ ~141 tok/s | ✓ ~82 tok/s | ✓ ~44 tok/s |
| Qwen3 32B | ✓ ~15 tok/s | ✓ ~8.6 tok/s | ✓ ~4.6 tok/s |
| Qwen2.5 Coder 32B | ✓ ~15 tok/s | ✓ ~8.6 tok/s | ✓ ~4.6 tok/s |
| DeepSeek R1 Distill 32B | ✓ ~15 tok/s | ✓ ~8.6 tok/s | ✓ ~4.6 tok/s |
| Mixtral 8x7B | ✓ ~38 tok/s | ✓ ~22 tok/s | ✗ |
| Llama 3.3 70B | ✓ ~7.1 tok/s | ✓ ~4.0 tok/s | ✗ |
| GLM-4.5-Air 106B | ✓ ~41 tok/s | ✗ | ✗ |
| gpt-oss-120b | ✓ ~94 tok/s | ✗ | ✗ |
| Mistral Large 123B | ✓ ~4.1 tok/s | ✗ | ✗ |
| Qwen3 235B-A22B | ✗ | ✗ | ✗ |
| GLM-4.5 355B | ✗ | ✗ | ✗ |
| Llama 3.1 405B | ✗ | ✗ | ✗ |
| 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.