What is KataGo Distributed Training?

KataGo Distributed Training (katagotraining.org) is the home of KataGo's first public-distributed training run. KataGo is a strong open-source, self-play-trained Go engine that can predict score and territory, play handicap games reasonably, and handle many board sizes and rules with the same neural net. The site coordinates volunteers who donate computing power to continue training from the end of KataGo's previous official run ("g170"), which ended in June 2020, and to see how much further the project can go.

What the site actually does

The core idea is distributed self-play. Instead of one organization running all the training games, many contributors run KataGo on their own machines, generate self-play games, and send the results back to a central server. Those games feed the training of new neural networks, which are then published back to the community.

The site's navigation reflects this loop:

  • Networks – the trained neural nets produced by the run
  • Games – self-play games generated by contributors
  • Contributions – records of who has contributed computing power
  • Extra Nets – additional networks beyond the main run
  • Archives – stored data from the run
  • Releases – KataGo engine releases
  • Stats for kata1 – statistics for the current main run

There are also links to the KataGo GitHub and the server GitHub, plus Log In / 注册 (register) options for contributors.

Why it exists

KataGo's earlier official training run stopped in June 2020. This site exists to resume that work through volunteer computing rather than a single centralized effort. The stated goal is to "see how much further we can go" — that is, to keep improving the engine's strength and capabilities beyond where the official run left off.

How the training has evolved

The Updates section shows the run is actively maintained and has moved through several technical milestones:

  • 2026-09-15 – Support for versions prior to v1.18.0 was retired. Training is described as "fast and healthy" with new transformer models, and community investigations suggest the transformers may be stronger than current training issues allow, with experiments underway to fix this.
  • 2026-08-25 – With v1.18.x released (adding transformer performance optimizations and rules fixes for data generation), the first transformers were uploaded to the run and self-play began switching to them.
  • 2026-07-29 – v1.17.0 released as the first official KataGo release supporting transformer models, described as stronger and more efficient than prior models. The main run planned to switch to transformers for self-play once enough contributors migrated.
  • 2026-04-08 – A new strong large model, "kata1-zhizi-b40c768nbt-fdx6c," was made available on the networks page and started being used for self-play, credited to hzyhhzy and ZhiziGo, with its training data to be mirrored on the KataGo data archive.
  • 2026-03-22 – A new strong model, "kata1-zhizi-b28c512nbt-muonfd2," was uploaded for rating games.

Who it's for

This site is relevant if you want to:

  • Use KataGo – download the engine and networks to analyze or play Go
  • Contribute computing power – join the distributed run and help generate self-play games (see the "How to Contribute" section on the site)
  • Follow progress – track training updates, new models, and run statistics

If you only want to run KataGo locally, you can use the releases and networks without contributing. If you want to help advance the engine, the contribution path is the main reason the site exists.

Practical notes

  • Contributing requires running a supported KataGo version. Older versions are progressively retired — as of the 2026-09-15 update, anything before v1.18.0 is no longer supported for contribution.
  • The run is community-driven, with named contributors and mirrored data archives.
  • The site does not present pricing information; it is an open-source, volunteer-based project.
katagotraining.org