What it does

Burn is a deep learning framework and tensor library written in Rust. It handles model training, inference, and numerical computing, with a focus on letting you write model code once and run it across a wide variety of hardware targets without rewriting.

The interesting bit

The entire framework is generic over a Backend trait, and backends are composable decorators. Wrap a GPU backend with Autodiff and it gains backpropagation; wrap it with Fusion and it gains automatic kernel fusion. There’s even a Remote backend (beta) that sends tensor operations over the network for distributed execution. This is the kind of type-system heavy design that Rust makes possible and Python frameworks can’t easily copy.

Key highlights

• Backend matrix covers Nvidia CUDA, AMD ROCm, Apple Metal, Vulkan, WebGPU, and CPU targets including WebAssembly and no_std embedded (via the Flex backend)
• Imports ONNX models and PyTorch/Safetensors weights, converting them to native Rust code that runs on any backend
• Includes a terminal training dashboard built on Ratatui for real-time metrics without external tools
• Same model code used for training and deployment — no translation layer or code changes required when moving from cloud GPU to browser WASM to embedded device

Caveats

• ONNX support is active development with a limited set of supported operators
• Only the Flex backend currently works in no_std environments
• WGPU backends can hit Rust’s default recursion limit during compilation; the README suggests manually raising it to 256

Verdict

Worth a serious look if you’re building ML systems where deployment targets vary wildly — cloud GPUs, user laptops, browsers, or embedded devices — and you want one codebase. If you’re already productive in PyTorch and only targeting Nvidia GPUs, the switch cost is harder to justify.