Skip to content

Cortiq 0.8.3

Choose a tag to compare

@infosave2007 infosave2007 released this 30 Sep 15:53
· 728 commits to master since this release

Whisper large-v3 and large-v3-turbo

  • Convert supported 128-bin Hugging Face Whisper checkpoints to CMF with cortiq convert.
  • Transcribe WAV PCM/float audio with cortiq transcribe on CPU, Apple Metal, or wgpu/Vulkan.
  • The mixed q4tp profile retains attention and the shared decoder token/output matrix in q8_2f, with feed-forward weights in q4tp.

Whisper weights are not included. Validate word error rate on your target audio before production use. See the README for conversion and transcription examples.

Validation

  • Workspace build and tests passed on Linux and macOS.
  • Clippy, CLI MSRV (Rust 1.88), and Vulkan-feature compile checks passed.
  • Release CLI built on the RTX PRO 4000 Linux pod; cortiq verify and English/Russian transcription smoke tests passed.