Candle

Candle is Hugging Face's machine-learning framework for Rust, built for people who want native performance without leaving a systems language. The API deliberately mirrors PyTorch (tensors, autograd, nn modules, optimizers), and it ships backends for CPU (MKL/Accelerate), CUDA, and Metal, plus WebGPU and WASM targets, so the same code can run on a server GPU, a Mac, or in the browser.

The examples directory is the real showcase: LLaMA, Mistral, Mixtral, Qwen, and Phi running as pure-Rust inference, plus Whisper for speech, Stable Diffusion for images, and Segment Anything for vision, all without a Python runtime. At 21,100+ stars and Apache-2.0 it is the default answer for shipping ML in Rust. The trade-off is ecosystem size: you get the core right, but the library of prebuilt models and training recipes is a fraction of PyTorch's, and CUDA builds still need the usual toolchain care.

1. Install Rust (rustup) and, for GPU builds, the CUDA toolkit with the right compute capability.
2. Add the crates: cargo add candle-core candle-nn (enable features like cuda, metal, mkl, or accelerate for your target).
3. Clone the repo and run an example: cargo run --example llama --release -- --model-id ... (see candle-examples).
4. For WASM, add the wasm32 target and build with the documented flags in the README.