Stockfish 19 brings stronger neural network, universal binaries and wider platform support

The Stockfish team has announced the release of Stockfish 19, a new major version of the world's leading open-source chess engine. The update is available as a drop-in replacement, promising stronger play and more accurate analysis in any compatible GUI. In test matches against its predecessor, Stockfish 18, the new engine gains up to 44 Elo points and wins more than three times as many game pairs as it loses, reinforcing the project's continued dominance atop engine championship standings.

At the heart of the upgrade is a revised neural network, SFNNv16, which trims the binary size by removing redundant threat features while adding new pawn-pair features for increased strength. The secondary neural network introduced back in Stockfish 16.1 has been retired altogether, improving performance in positions where the smaller net had previously underperformed. The training pipeline also saw refinements, including Quantization-Aware Training (QAT) and parameter tweaks, applied across hundreds of billions of positions rescored with a strong Leela network.

Beyond raw playing strength, version 19 focuses on usability and reach. The project has moved to universal binaries that automatically detect a CPU's capabilities, sparing users from manually choosing between instruction sets such as AVX2 or AVX-512. Native support has been added for RISC-V and LoongArch processors, alongside 1 GB Linux huge pages and WebAssembly targets, and position validation is now stricter, with the engine terminating immediately and reporting a “CRITICAL ERROR” when fed invalid input.