burn Releases

Burn 0.21.0 brings 4 months of improvements focused on speed and reliability across distributed workflows and inference. Key updates include a rethought distributed computing stack using differentiable collective operations, a new lightweight `burn-flex` CPU backend for WebAssembly, and `burn-dispatch` for simplified backend selection. The release also adds early off-policy reinforcement learning support and kernel optimizations for GEMV, top-k, and FFT.

May 7, 2026

This release focuses on bug fixes and performance improvements. It addresses a RAM memory leak in autodiff, improves memory reuse with a sliced memory pool, and updates CUDA support to fix 12.8 features and auto-detect the version.

May 7, 2026

This release focuses on stabilizing the system with bug fixes and usability improvements, specifically targeting non-CPU device support, voice cloning stability, and adding language-specific default voices.

May 4, 2026

This release enhances Difftastic's ability to ignore insignificant trailing punctuation in multiple languages and adds support for Assembly parsing. It also improves parsers for Clojure, Common Lisp, CSS, Dart, Erlang, F#, OCaml, Python, Rust, Scala, TypeScript, and VHDL.

Apr 30, 2026

A major version update introducing extensive multilingual support and significant model improvements, alongside performance optimizations and architectural changes.

Apr 21, 2026

This update resolves a critical Git integration issue where difftastic would crash on unmerged paths and improves error reporting for invalid arguments. It also updates parsers for Bash, C, Go, Lua, Nix, Perl, Python, Rust, Scala, Swift, and YAML.

Mar 16, 2026

This maintenance release prioritizes documentation improvements and adds CUDA generation support to enhance compatibility and usability.

Feb 16, 2026

A feature-rich release focused on developer experience, including online documentation, local model loading, state export/import functionality, and CLI improvements.

Feb 16, 2026

A major release enabling efficient distributed training, quantization, and CPU support. It introduces multi-stream execution for multi-GPU parallelism, a `burn-collective` crate for gradient synchronization, and a new CPU backend powered by MLIR and LLVM. The training loop was refactored to support distributed strategies, and comprehensive quantization support was added.

Jan 29, 2026

Patch release with bug fixes and improvements for the Burn deep learning framework. Fixes include book guide training changes and dequantize native debug statements. Also removes pinned exact versions to allow pulling patch releases.

Jan 23, 2026

A bug fix release addressing Windows compatibility issues, audio generation parameter handling, and providing CLI control over chunk sizes.

Jan 22, 2026

Major release introducing CubeK for unified CPU and GPU kernels through CubeCL, achieving peak performance across diverse hardware. Features complete ONNX import system overhaul for broader model support, new tensor operations, and significant performance improvements.

Jan 15, 2026

This release marks a significant step forward in performance and reliability. It features optimized matrix multiplication kernels specialized for mat@vec, vec@mat, inner, and outer products. Fusion was enhanced with dead code elimination and a search engine, and critical multi-threading issues were resolved with new concurrent stream support.

Dec 17, 2025

This release introduces hyperparameter tuning, data synthesis capabilities, and AWS Bedrock integration, alongside enhanced launcher functionality for HPC clusters.

Nov 18, 2025

This release introduces support for Protocol Buffer files and significantly improves parsing for Ada, Clojure, CMake, and several other languages. It also optimizes the display logic for side-by-side diffs and binary file descriptions.

Nov 16, 2025

Nov 9, 2025

A bug-fix focused release addressing memory leaks, CUDA compatibility, and Tensor API regressions to improve stability and performance.

Nov 6, 2025

A major overhaul introducing multi-stream execution for distributed training, comprehensive quantization support, and a new CPU backend powered by MLIR and LLVM.

Oct 28, 2025

Adds support for Qwen3-VL models and NERSC Perlmutter HPC clusters, alongside enhancements to logging and performance.

Oct 20, 2025

This release focuses on build optimizations for systems with large page sizes (Arm) and introduces stricter text encoding detection to prevent binary files from being mishandled. It adds a new command-line option to override binary file detection.

Sep 24, 2025