Burn is a next-generation deep learning framework written in Rust, built as both a tensor library and a complete training and inference framework. Its central idea is to end the train-then-export split: with most stacks you train in Python, export to ONNX and hope nothing is lost in translation to a production engine, whereas Burn runs the exact same code in training and in production. It keeps the intuitive ergonomics of PyTorch, with dynamic shapes and graphs, while JIT-compiling tensor operations for speed. Burn has more than 16,000 stars on GitHub and is actively developed in 2026.
Burn executes multi-platform tensor operations through a single unified API, with backends for CUDA, ROCm, Metal, Vulkan, WebGPU and CPU, so models can train on a workstation and deploy to servers, desktops, browsers or edge devices without being rewritten. Kernel fusion and autotuning keep performance competitive, and ONNX import lets you bring existing models in. Because deployment is native rather than an export step, advanced scenarios like on-device personalization and federated learning become straightforward. The framework is developed by Tracel AI and is aimed at teams who want Rust’s safety and performance across the whole model lifecycle.
Burn is free and open source under the Apache-2.0 and MIT dual license. There are no paid tiers; you run it on your own hardware.
Burn is for ML engineers and Rust developers who want one framework from prototype to production, especially for deploying models to edge devices, browsers and other targets where Python-based stacks are awkward.