Hacker News new | ask | show | jobs
by storus 246 days ago
M5 is supposed to support FP4 natively which would explain the speed up on Q4 quantized models (down from BF16).