Uzu is a high-performance AI inference engine written in Rust, designed for low-latency, private, and cost-free local model deployment. The engine supports multiple backends, including Metal, WebGPU, and CPU, with cross-platform compatibility spanning macOS (aarch64 and x86_64), iOS, and ongoing support for Windows, Linux, and WebAssembly (WASM). It provides high-level language bindings for Rust, Python, Swift, and TypeScript, enabling developers to manage the full lifecycle of AI interactions—from automated model downloading to complex inference. Key technical capabilities include structured output generation via JSON Schema grammar, tool-calling functionality using decorated functions, and support for reasoning-effort configurations. To optimize memory usage, particularly on Apple devices through unified memory, the engine allows for shared chat instances that reuse loaded model weights across multiple sessions. Additionally, uzu can function as an OpenAI-compatible HTTP server, exposing endpoints such as /v1/chat/completions for standard chat completions and streaming. The ecosystem is supported by "cargo tools" for streamlined dependency management and "lalamo" for converting models into the engine's native format.