mimi-rs

Rust implementation of the Mimi audio codec by Kyutai Labs, built on candle.

Shared library used by tts-web and stt-web for streaming audio encoding/decoding in the browser via WASM.

Components

SEANet encoder/decoder for audio compression
Streaming transformer with KV cache (Mimi-style context window) and growing KV (FlowLM-style)
QLinear for post-load F32 → Q8_0 weight quantization via candle's QMatMul
Rotary embeddings, layer scaling, residual convolutions, resampling

[dependencies]
mimi-rs = { git = "https://github.com/idle-intelligence/mimi-rs.git" }

Model weights: Kyutai license. Code: MIT.