KataGo

KataGo is an open-source computer program that plays the board game Go at a very high level. As of 2026, it is described as one of the strongest open-source Go bots available. It was built using a self-play training approach: the program plays millions of games against itself and uses the results to improve, without needing any human game records as a starting point.

The training method is based on techniques similar to those used in DeepMind's AlphaZero, but with a number of improvements that make it substantially faster to train. According to the README, a researcher with a few powerful graphics cards and a few days of compute time could train a bot from scratch to a high amateur level on the full 19x19 board. With a single high-end consumer GPU and several months of time, it may even be possible to train to superhuman strength.

Beyond raw strength, KataGo is designed as a practical analysis tool for Go players. It estimates territory and score rather than just win probability, which makes it useful for reviewing games at all skill levels, not only professional play. It supports board sizes from 7x7 to 19x19, various rulesets including Japanese rules, different komi values, and handicap games. For developers, it includes a JSON-based analysis engine that can process multiple board positions in batches.

KataGo itself has no graphical interface. It communicates through a standard text protocol called GTP, which means it is meant to be paired with a separate GUI application. The README lists several options, including KaTrain, which bundles KataGo and is described as the easiest choice for non-technical users, as well as Sabaki and others for those who want more control.

Precompiled executables for Windows and Linux are available on the releases page. The latest trained neural network files are hosted separately at katagotraining.org, where an ongoing public distributed training run is also taking place.