oMLX

DFlash Doubles the T/S Gen Speed of Qwen3.5 27B (BF16) on Mac M5 Max

The new DFlash support in oMLX 0.3.5 RC1 has reportedly doubled the generation speed of the Qwen3.5 27B (BF16) model on a Mac M5 Max, increasing it from 9 to 22 T/S. This breakthrough could significantly improve local deployment of this high-quality model at higher quantizations/full weights.