KataGo

open_source

KataGo is a free, open-source Go-playing AI engine using self-play reinforcement learning. Features distributed training, multi-board-size support, and score prediction.

AI Models & Infrastructure

Foundation Models

AI Research Tools

About

KataGo is one of the strongest open-source Go engines available, built using deep reinforcement learning and self-play training techniques. Originally developed with novel algorithmic improvements documented in an arXiv paper, KataGo continues to evolve through a community-driven distributed training program hosted at katagotraining.org. The engine features a versatile neural network architecture capable of handling many board sizes (including experimental support up to 50x50) and multiple Go rulesets within a single model. It can predict final scores, assess territory, and play handicap games effectively. Volunteers worldwide contribute GPU compute to help train progressively stronger neural networks, with multiple model sizes available — from lighter b18 networks to large b28 and b40 models for maximum strength. KataGo is widely used by Go players wanting analysis tools, developers building Go applications, and AI researchers studying self-play learning techniques. It integrates with popular Go software like Sabaki, Lizzie, and KaTrain via the GTP protocol. The project regularly releases new versions with performance improvements, new training data formats, and expanded research capabilities such as experimental action-value heads. All networks, training data, and source code are freely available under open-source licenses, making KataGo an ideal platform for both practical Go analysis and cutting-edge AI research.

Key Features

Self-Play Reinforcement Learning: KataGo trains entirely through self-play, continuously improving its neural networks without human game data, achieving superhuman Go strength.
Distributed Community Training: Volunteers worldwide contribute GPU compute to the public distributed training run, collaboratively pushing the model's strength further over time.
Score & Territory Prediction: Beyond just move selection, KataGo predicts final scores and territory ownership, making it an excellent analysis and review tool for Go players.
Multi-Board Size & Rules Support: A single neural network handles many board sizes (up to 50×50 experimentally) and multiple rulesets, offering broad compatibility across Go variants.
Multiple Network Sizes: Offers a range of neural network sizes (b18, b28, b40) to balance speed and strength, suitable for everything from fast analysis to maximum-strength play.

Use Cases

Go players using KataGo as an analysis engine to review games, identify mistakes, and study optimal moves at a superhuman level.
AI researchers studying self-play reinforcement learning techniques using KataGo's open training data, code, and published methodology.
Developers building Go applications, bots, or training platforms that integrate KataGo via its GTP protocol interface.
Volunteers contributing GPU compute to the distributed training run to help advance open-source game AI research.
Educators and students exploring deep reinforcement learning concepts through a real-world, well-documented open-source project.

Pros

Completely Free & Open Source: All source code, trained networks, and training data are freely available, making KataGo accessible to everyone from hobbyists to professional researchers.
World-Class Go Strength: KataGo consistently ranks among the strongest Go AI engines available, offering superhuman analysis capabilities for players at all levels.
Active Community & Ongoing Development: A thriving community of contributors, researchers, and developers ensures regular releases, new networks, and continued strength improvements.

Cons

Technical Setup Required: Running KataGo locally requires command-line knowledge and configuration; it is not a plug-and-play consumer application out of the box.
Narrow Domain Focus: KataGo is purpose-built for the game of Go and has no utility outside of Go analysis, play, or related AI research.
Hardware Demands for Strongest Models: The largest and strongest networks (b28, b40) require significant GPU resources to run at full speed, which may be a barrier for some users.

Frequently Asked Questions

KataGo is an open-source Go-playing AI engine trained via self-play reinforcement learning. It is one of the strongest publicly available Go engines and supports features like score prediction, territory analysis, multiple board sizes, and various rulesets.

Yes. KataGo is completely free and open-source. All neural network weights, training data, and source code are publicly available at no cost.

You can contribute by running KataGo's contribute mode on your machine, which uses your GPU to generate self-play games that are uploaded to the distributed training server, helping train stronger future networks.

KataGo supports the GTP (Go Text Protocol), making it compatible with popular Go interfaces such as Sabaki, Lizzie, KaTrain, GoGui, and many others.

For maximum strength, the latest b28 or b40 series networks are recommended if your hardware can support them. For faster analysis on modest hardware, the b18 networks offer an excellent speed-to-strength tradeoff.