Models & Hardware
MiniMax M3.1 Flash Runs on a Mac at 42.0 Tok/s With oMLX
The bundled DSpark draft lifts one request on an M3 Ultra from 22.5 tok/s, an 87% gain, @jundotkim says.
@jundotkim says MiniMax M3.1 Flash now runs on Macs with oMLX 0.7.3.dev1.
The gain comes from the model’s bundled DSpark draft, running as Lightning MTP. One request on an M3 Ultra with the oQ4e quant goes from 22.5 to 42.0 tok/s (+87%). New fused decode kernels add another 6-7%. Prefill runs at 370 tok/s at 4K and still 298 tok/s at 128K.
| Quant | Size |
|---|---|
| oQ3e | 196GB |
| oQ3.5e | 213GB |
| oQ4e | 249GB |
The release speeds up other models too. Qwen3.8-27B fp16 with DFlash2 on an M3 Ultra goes from 106 to 141 tok/s. Two GLM-5.3 Flash requests at 64K go from 16.4 to 19.2 tok/s each, and the model uses 3.27 GB less memory.
Sources
- @jundotkim on X · Oct 24, 9:55 AM PT
- oMLX 0.7.3.dev1 release notes · github.com