Tools
whisper.cpp 1.9.4 pulls in Metal 4.0 tensor API support for M5 and A19 chips and sparse flash attention
The 2026-09-11 release syncs ggml 0.23.0 and carries `metal: enable Metal 4.0 tensor API on M5+/A19+`, `metal: add sparse FA`, and flash-attention-vec tunings for M2 Pro, M2 Max, M3, M3 Max and A18 Pro. It also adds Windows on ARM to the release job and fixes several missing autorelease pools that leaked memory on Metal. On the whisper side, the encoder_begin_callback now fires before language auto-detect, and the server returns the detected language in its detect response.
Source
↳ Follow the thread