burn

Unlike Python-first frameworks that rely on native extensions, Burn is built entirely in Rust from the ground up, enabling it to compile to native code, WebAssembly, and even run on embedded devices without a Python runtime dependency.

The framework supports multiple backends interchangeably Candle (pure Rust), LibTorch (PyTorch C++ bindings), WGPU (cross-platform GPU compute via WebGPU), and NdArray (CPU baseline) letting researchers prototype on CPU and deploy to GPU without code changes.

Burn's architecture centers on a backend-agnostic tensor API where neural network modules, optimizers, and loss functions are written once and execute on any supported backend.

This design enables true hardware portability: the same model definition runs on NVIDIA GPUs via CUDA, Apple Silicon via Metal, AMD GPUs via ROCm, and WebGPU in browsers.

The framework includes built-in support for automatic differentiation, mixed precision training, gradient checkpointing, and data loading pipelines, covering the essential infrastructure for training modern neural networks.

The Rust ML ecosystem is growing rapidly, and Burn positions itself as the production-grade framework for teams that need performance guarantees, memory safety, and deployment flexibility that Python frameworks cannot easily provide.

Applications range from training custom models in compute-intensive research pipelines to deploying inference at the edge on IoT devices and in WebAssembly environments.