kimi-k3-in-rust
A byte-identical Rust port of kimi-k3-in-c that runs CPU-only inference for the 2.78-trillion-parameter Kimi K3 mixture-of-experts model, with no GPU, BLAS or PyTorch, in an 8 GB memory floor.
kimi-k3-in-rust is a statement-for-statement Rust port of the C99 kimi-k3-in-c engine, including its CLI, tests and benchmark, for CPU-only inference of the 2.78 trillion parameter Kimi K3 mixture-of-experts model. It uses four dependencies and no BLAS, framework or GPU, streams the checkpoint from disk, and holds to the same 8 GB memory floor as the C original.
A test suite confirms the Rust engine's logits are byte-identical to the C build on both tiny models and the full released checkpoint, and on real weights the Rust port decodes 1.54x to 1.62x faster per token than the C build on aarch64, with continuous integration checking Linux x86-64, macOS arm64, Clippy and formatting.
Built with Rust, using only libc, rayon, serde and serde_json as dependencies. Apache-2.0 licensed.