Hi MOSS Team and Community! 👋
First off, huge thanks for open-sourcing MOSS-TTS-Nano! The model quality, architecture elegance, and the paper itself have been incredible resources.
I'm an independent developer who was inspired to build a high-performance Rust port of your work using the Hugging Face Candle framework. The result: numerical parity with the Python reference, a ~10 MB static binary, and ~0.2s startup time (vs ~5s Python).
🔗 Rust Port Repo: https://github.com/ramishi/moss-tts-nano-rust-candle
Key Features
✅ Voice cloning & continuation mode (parity verified)
✅ Stereo 48kHz Float32 WAV output
✅ Zero Python dependency (single binary deployment)
✅ HuggingFace auto-download + offline mode
✅ Full sampling controls (temp, top-p, seed, etc.)
🚧 CUDA/MPS GPU support (planned)
🚧 Streaming API (planned)
Why This Might Interest You
Embedded/Edge deployment: No Python runtime needed
Production API servers: Rust's async ecosystem (axum/actix)
Research: Easy to extend/modify the Rust codebase
Performance: ~0.07x RTF on Apple M4 (CPU)
Call for Feedback
I'd love your thoughts on:
Numerical parity: I've verified tensor-level match, but more edge cases?
Architecture: Any Rust-specific optimizations you'd suggest?
Features: What should be prioritized next? (CUDA? Streaming? ONNX export?)
Collaboration: Would the team be open to cross-pollination or mentions in docs?
The repo includes 30 passing tests, Clippy-clean code, and a GitHub Actions CI pipeline. If you have any feedback, issues, or want to contribute, please visit the repo or drop a comment here!
Again, massive kudos to @YitianGong, @botianjiang, and the entire OpenMOSS team. This project wouldn't exist without your pioneering work.
Hi MOSS Team and Community! 👋
First off, huge thanks for open-sourcing MOSS-TTS-Nano! The model quality, architecture elegance, and the paper itself have been incredible resources.
I'm an independent developer who was inspired to build a high-performance Rust port of your work using the Hugging Face Candle framework. The result: numerical parity with the Python reference, a ~10 MB static binary, and ~0.2s startup time (vs ~5s Python).
🔗 Rust Port Repo: https://github.com/ramishi/moss-tts-nano-rust-candle
Key Features
✅ Voice cloning & continuation mode (parity verified)
✅ Stereo 48kHz Float32 WAV output
✅ Zero Python dependency (single binary deployment)
✅ HuggingFace auto-download + offline mode
✅ Full sampling controls (temp, top-p, seed, etc.)
🚧 CUDA/MPS GPU support (planned)
🚧 Streaming API (planned)
Why This Might Interest You
Embedded/Edge deployment: No Python runtime needed
Production API servers: Rust's async ecosystem (axum/actix)
Research: Easy to extend/modify the Rust codebase
Performance: ~0.07x RTF on Apple M4 (CPU)
Call for Feedback
I'd love your thoughts on:
Numerical parity: I've verified tensor-level match, but more edge cases?
Architecture: Any Rust-specific optimizations you'd suggest?
Features: What should be prioritized next? (CUDA? Streaming? ONNX export?)
Collaboration: Would the team be open to cross-pollination or mentions in docs?
The repo includes 30 passing tests, Clippy-clean code, and a GitHub Actions CI pipeline. If you have any feedback, issues, or want to contribute, please visit the repo or drop a comment here!
Again, massive kudos to @YitianGong, @botianjiang, and the entire OpenMOSS team. This project wouldn't exist without your pioneering work.