oMLX on the M3 Ultra 512GB
On the M3 Ultra 512GB, oMLX runs DeepSeek V4 Flash 284B · 13B active at 8-bit at a median 20.0 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 | oMLX | Other engines |
|---|---|---|
| DeepSeek V4 Flash 284B · 13B active 8-bit | 20.0 t/sone report, context not stated | no other engine reported |
| Qwen3.8 125B · 6B active 4-bit | 25.0 t/sone report, context not stated | llama.cpp 31.1 t/sone report, 250K context |
4 oMLX 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.
- DeepSeek V4 Flash 284B · 13B active · MXFP414.9 t/s(speculative decoding)
- Qwen3.8 125B · 6B active · 4bit25.0 t/s
- DeepSeek V4.1 Flash 552B · 16B active · oQ4e19.7 t/s(part in system RAM)
- DeepSeek V4 Flash 284B · 13B active · 8bit affine20.0 t/s
Frequently asked
How fast is oMLX on the M3 Ultra 512GB?
On the M3 Ultra 512GB, oMLX runs DeepSeek V4 Flash 284B · 13B active at 8-bit at a median 20.0 tokens per second for generation, from 1 plain run (context not stated).
Is oMLX faster than other engines on the M3 Ultra 512GB?
Same card, same model size and bit class, plain runs only: Qwen3.8 125B · 6B active at 4-bit gives oMLX 25.0 t/s (1 run) and llama.cpp 31.1 t/s (1). Context length and engine build differ between the runs, so treat it as a lead, not a benchmark.