Experimental local voice runtime for Cherry Studio. This is a downloadable runtime/model artifact release, not an application release.
Platform: Apple Silicon, macOS 26.2 or newer. Users do not need Python, Rust, Xcode, or a separate Node installation.
- mlx-node source: https://github.com/mlx-node/mlx-node/tree/5db3d24bc87a9eb6fdb48c7ab3c9d8f126bf8d7d
- MLX source: https://github.com/ml-explore/mlx/tree/e091d678454e2462efbad8b45b87a3948bb05b38
- Model: https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice/tree/85e237c12c027371202489a0ec509ded67b5e4b5
- Quantization: affine 8-bit, group size 64; dense speech codec.
- Runtime/model checksums and lengths: manifest.json. Runtime MIT licenses are included in the archive; model Apache-2.0 license is attached.
- Build/packaging instructions and script are attached. Artifacts must not be replaced in place; changes require a new tag and checksum manifest.
Validated on Apple M5, macOS 26.5, Electron 44.4.4: checksum-verified installation through Cherry's installer, utility-process synthesis, multilingual voice selection, finite WAV playback, cancellation during generation and playback, and reuse after cancellation. The app currently uses complete WAV output through the shared player; native streaming does not imply streaming playback in Cherry. Other OS/architecture combinations are not supported by this artifact.