Tract is a lightweight, self-contained inference runtime written in Rust that supports ONNX and NNEF model formats. It optimizes and compiles neural-network models for various backends including embedded ARM CPUs, NVIDIA/Apple GPUs, and browser WebAssembly. The project is used in production at Sonos for wake-word detection and streaming speech-recognition, and supports running LLMs, text-to-image, and computer-vision models.