What Are the Key Features of the KataGo Go Engine?
KataGo is a strong open-source Go engine trained through self-play, and its main distinguishing feature is that a single neural net handles score and territory prediction, reasonable handicap play, and many board sizes and rules at once. It also incorporates learning-acceleration techniques described in its arXiv paper and later work, and current releases support transformer models that are stronger and more efficient than earlier architectures. These features matter if you want one engine that adapts across rulesets and board sizes rather than a specialized tool per setting.
Core Capabilities
The project describes KataGo as "a strong open-source self-play-trained Go engine, with many improvements to accelerate learning (arXiv paper and further techniques since)." From that description, the practical capabilities are:
- Score and territory prediction — the engine can predict both the score and the territory, not just suggest moves.
- Handicap games — it "play[s] handicap games reasonably," which many engines handle poorly.
- Multiple board sizes and rules — it handles "many board sizes and rules all with the same neural net," so you don't need separate models per configuration.
- Self-play training — strength comes from self-play rather than only human game data.
The phrase "all with the same neural net" is the key design point: board size and rules are inputs to one model, not reasons to switch models.
Transformer Models
Recent releases changed the model architecture. According to the site's update log:
- v1.17.0 (2026-07-29) was "the first official release of KataGo that supports transformer models, which are stronger and more efficient than all the prior models."
- v1.18.x (2026-08-25) added "additional performance optimizations for transformers and some rules fixes for data generation," and the first transformers were uploaded to the distributed training run.
- 2026-09-15 — support for versions prior to v1.18.0 was retired.
So if you are choosing a version today, the transformer-capable line (v1.18.0 or later) is the one the project supports for contribution, and the one associated with stronger, more efficient models.
What This Means in Practice
| Feature | Why it matters to you |
|---|---|
| One net for many board sizes/rules | Run 9x9, 19x19, and other configurations without swapping models |
| Score and territory prediction | Useful for analysis and review, not just move suggestions |
| Reasonable handicap play | Better behavior in uneven games |
| Self-play training | Strength improves through the project's training runs |
| Transformer models (v1.17.0+) | Stronger and more efficient than prior architectures |
Where the Distributed Training Fits
KataGo's engine is tied to an ongoing public-distributed training run hosted at katagotraining.org. The site states it "hosts KataGo's first public-distributed training run," attempting to resume training from the end of the previous official run ("g170") that ended in June 2020. Volunteers contribute computing power, and the resulting networks are published for download. This is why the engine's strength continues to change over time — the models you download depend on how far the current run has progressed.
If your goal is simply to use the engine, the relevant takeaway is that you should download a current network alongside a supported KataGo release (v1.18.0 or later), since older versions are no longer supported for contribution and newer models are built around the transformer architecture.