LeelaChessZero/lc0

Lc0 is a UCI-compliant chess engine that plays chess using neural networks trained by the LeelaChessZero project.

The engine solves the problem of creating a strong chess-playing AI by implementing neural network evaluation rather than traditional hand-crafted evaluation functions. It works by accepting neural network weights from the LeelaChessZero training infrastructure and using them to evaluate board positions during search. The engine supports GPU acceleration across multiple platforms and can also run on CPU using BLAS libraries, making it accessible to users with different hardware configurations.

Developers should choose this tool if they want to run or contribute to neural network-based chess AI. It suits projects involving competitive chess play, self-play game generation for training, or research into deep reinforcement learning applied to chess. The project is tightly integrated with the LeelaChessZero training server, which enforces game quality standards based on engine versions, so users participating in the distributed training effort should use release branches rather than development versions.

Development activity shows consistent attention to build infrastructure, with the README emphasizing that building has become easier and encouraging users to report problems. The project maintains broad hardware support across NVIDIA GPUs via CUDA, AMD GPUs via HIP and ROCm, Intel GPUs via experimental SYCL, and Apple hardware via Metal Performance Shaders, alongside CPU backends using OpenBLAS or DNNL. The codebase requires C++20 support and uses the Meson build system with Ninja, reflecting a modern development approach. The project follows semantic versioning and distinguishes between release branches for stable play and development branches for experimental work.