MLX on the M3 Ultra 512GB
On the M3 Ultra 512GB, MLX runs DeepSeek V4 Flash 284B · 13B active at 4-bit at a median 29.9 tokens per second for generation, from 1 plain run (context not stated).
Speed by model on the M3 Ultra 512GB
Plain runs only: one device, one request, no speculative decoding, the whole model in memory. Each row is one model size at one bit class, with the other engines' plain runs of the same model on this card beside it. Context length and build still differ between runs, and each side shows its own.
| Model | MLX | Other engines |
|---|---|---|
| DeepSeek V4 Flash 284B · 13B active 4-bit | 29.9 t/sone report, context not stated | no other engine reported |
| Llama 3.1 405B 4-bit | 3.0 t/sone report, 1K context | no other engine reported |
| Qwen3 32B 4-bit | 31.2 t/sone report, 1K context | no other engine reported |
15 MLX reports on the M3 Ultra 512GB
Newest first. A run on several cards, with speculative decoding, with part of the model in system RAM or with batched requests says so, since each describes a different setup.
- Llama 3.1 · 4-bit14.9 t/s
- Qwen2.5 · 4-bit14.3 t/s
- Llama 3.3 · Q86.8 t/s
- Gemma 4 · BF169.2 t/s
- Qwen3.8 · Q431.8 t/s
- Qwen3.6 · Q489.3 t/s
- Qwen3.5 · Q449.5 t/s
- Qwen3 · Q435.8 t/s
- gpt-oss 117B · 5.1B active · MXFP4/BF16 MLX62.6 t/s
- Qwen3.8 · MLX affine 4-bit, group 64, 4.501 effective bpw37.9 t/s
- Llama 3.1 405B · Q43.0 t/s
- DeepSeek V4 Flash 284B · 13B active · 4-bitPP 1012.0 t/s(2 cards, speculative decoding)
- DeepSeek V4 Flash 284B · 13B active · 4bit29.9 t/s
- Kimi K3 2800B · 104B active · MXFP42.0 t/s(4 cards)
- Qwen3 32B · Q431.2 t/s
Frequently asked
How fast is MLX on the M3 Ultra 512GB?
On the M3 Ultra 512GB, MLX runs DeepSeek V4 Flash 284B · 13B active at 4-bit at a median 29.9 tokens per second for generation, from 1 plain run (context not stated).