Visit website
Updated: 2026-10-08 11:42
Language: English (default)
Access: Normal
Profile views 1 
Outbound visits 0 Website Review
What is KataGo Distributed Training?
KataGo Distributed Training is a volunteer-powered project that continues training KataGo, a strong open-source Go engine, using distributed self-play. Rather than one lab running all the games, contributors donate computing time; their self-play games feed a shared training run that produces updated neural networks. The site describes this as KataGo's first public-distributed training run, resuming from the end of the earlier official "g170" run that stopped in June 2020, to see how much further the engine can improve.
What it actually does
- Crowdsources self-play data. Volunteers run KataGo locally to generate games, which are submitted to the central server. That data trains new network versions.
- Publishes networks and releases. The site has pages for downloading KataGo, browsing networks, and viewing archives, plus links to the KataGo engine and server code on GitHub.
- Tracks progress publicly. A stats section for the current run ("kata1") and an updates log show training activity and milestones.
Who it's for
- Go players and engine users who want the latest, strongest KataGo networks for analysis or play.
- Volunteers with spare GPU/CPU capacity willing to run self-play and contribute games.
- Researchers and developers interested in self-play training, transformer models, and distributed compute.
Practical trade-offs
Contributing means installing and maintaining a recent KataGo version, since the project retires support for older versions as the run moves forward — for example, the updates note that versions before v1.18.0 were retired, and that the run switched to transformer models for self-play. That keeps data quality and efficiency high, but it puts an upgrade burden on contributors. Downloading networks is simpler: you get stronger models without contributing compute, though you depend on the project's release cadence.
Where to start
If you only want the engine, go to the download section and pick a current release. If you want to help train it, read the "How to Contribute" section and make sure your version is current before running self-play. For the underlying engine and server software, see GitHub.
How do I contribute my computer to the KataGo distributed training run?
To contribute your computer, you install a recent KataGo release, run its self-play generation client, and connect it to the public distributed run so your machine plays games and uploads the resulting data. The project describes this as a volunteer effort to continue training from the end of its earlier official run.
What you are actually donating
Your CPU or GPU generates self-play games with the current network. Those games become training data, and the run uses them to produce the next network. This is a good fit if you have a reasonably modern GPU, are comfortable running a command-line program for long stretches, and want to support Go engine research without running your own training pipeline.
It is a poor fit if you need guaranteed uptime, want to use the machine interactively at full speed, or dislike background network usage. Generation is compute-heavy and sustained; a laptop on battery or a shared work machine is usually the wrong host.
Practical steps
- Download KataGo from the official site KataGo Distributed Training or the project's GitHub releases, and check the site's "Downloading KataGo" section for the current recommended version.
- Register or log in on the training site, since contributions are tied to an account.
- Follow the "How to Contribute" instructions to configure and launch the self-play client with your credentials.
- Leave it running; the site's stats for the current run show whether your games are being accepted.
- Keep the version current. The run's own updates note that support for versions before v1.18.0 was retired, and earlier notes warned that older versions would be disabled once enough contributors migrated.
Version matters more than raw hardware
The update log is unusually explicit that the run changes underneath you. Transformer models were introduced with v1.17.0 and then adopted for self-play in the main run; v1.18.x added further transformer performance work and rules fixes for data generation. An outdated client can therefore stop being accepted, or contribute less useful data, even if it still runs.
| Situation | Sensible action |
|---|---|
| New contributor | Install the latest release first, then register and configure |
| Already contributing on an older build | Upgrade before the next support cutoff |
| Limited GPU time | Contribute fewer hours consistently rather than chasing maximum throughput |
| Machine also used for work | Cap threads or GPU usage, or contribute only when idle |
As a concrete example: someone with a single mid-range GPU might run the client overnight and on weekends, confirm on the stats page that games are landing, and upgrade whenever the site announces a new minimum version. That produces steady, useful data without turning the desktop into an unusable machine.
Your next step is to open the site's "How to Contribute" section, confirm the minimum supported version, and complete registration before starting the client. If you would rather just play or analyze with KataGo, downloading the engine and a network is the better route; distributed training is for people willing to keep a client running and maintained.
What is the difference between the g170 run and the current kata1 distributed run?
The g170 run was KataGo's previous official self-play training run, which ended in June 2020. The kata1 run hosted on this site is a volunteer-driven effort to resume training from where g170 stopped and see how much further the project can go. In other words, kata1 is a continuation rather than a fresh start, but it is organized as a public distributed run rather than an official in-house one.
What changed in practice
- Who supplies the compute: g170 was the official run; kata1 relies on volunteers contributing self-play games and training resources through the site's contribution workflow.
- Model generation: kata1 has moved through several model generations, including large models contributed by community members and, more recently, transformer-based models that are described as stronger and more efficient than prior architectures.
- Software requirements: contributing to kata1 now requires recent KataGo versions. Older versions have been retired, and support for versions before v1.18.0 has been dropped.
- Ongoing tuning: the run is still being actively investigated and adjusted, with reported training issues in the transformer models being worked on for further improvement.
How to decide which matters to you
If you want to play or analyze with the strongest current networks, use the latest kata1 models and a recent KataGo release. If you want to contribute games, check the site's contribution instructions and confirm your client version is current, since outdated versions are no longer accepted. For background on the engine itself, see GitHub (KataGo).
Which KataGo version do I need to download to participate in distributed training?
You need KataGo v1.18.0 or later to contribute to the distributed training run at KataGo Distributed Training. The run's own update notes state that support for versions prior to v1.18.0 has been retired, and that the server may disable older-version contributions once enough volunteers have migrated.
Why the cutoff exists
The current run has switched to transformer models, which are stronger and more efficient than earlier architectures. v1.17.0 was the first official release to support transformers, but v1.18.x added further performance optimizations for them plus rules fixes for data generation. Since the training pipeline now generates and consumes self-play data in the transformer format, older clients no longer produce usable contributions.
Practical steps
- Go to the official KataGo releases page and download the latest v1.18.x build for your platform (Windows, Linux, macOS, or via a package manager).
- Download a current transformer network from the networks page, since the older non-transformer networks are no longer what the run trains on.
- Run the client in distributed-training mode, following the "How to Contribute" section of the site.
- Keep the client reasonably up to date. The run's history shows version requirements tightening over a few weeks after each release, so an outdated build can silently stop earning contributions.
Trade-off to weigh
Upgrading costs you time and possibly a re-tune of your configuration, and transformer networks are larger and often need a stronger GPU to hit good throughput. The payoff is that your games actually count toward the run, and transformer self-play is reported to be faster and healthier than the previous setup. If your hardware is old or low-end, check the release notes for CPU/GPU requirements before committing, because a slow client may contribute little even if it is version-compatible.
What are transformer models and why is KataGo switching to them?
Transformer models are a newer neural-network architecture that KataGo's distributed training run has adopted to replace its earlier convolutional networks. In practice, switching architecture changes how the engine learns from self-play: the project's own updates describe transformers as stronger and more efficient than prior models, and the run has begun uploading transformer networks for self-play rather than only for evaluation.
What the switch means for users
- Strength and efficiency: KataGo's release notes state that transformer models outperform earlier models, so the main training run is migrating to them.
- Version requirement: Transformer support arrived in v1.17.0 and was extended in v1.18.x. Older clients are being retired; the run's update notes say support for versions before v1.18.0 has been retired, so contributing with an old client will stop working.
- Ongoing tuning: The project notes that recent investigations suggest transformers may be underperforming their potential because of training issues, and experiments are underway to fix this.
Why switch now
The distributed run depends on volunteers generating self-play games. If contributors stay on old versions, the run cannot use transformer networks consistently, so the project is phasing out older clients once enough people upgrade. For a volunteer, the practical step is to upgrade to the latest KataGo release and confirm your client is generating games with the current network. For a user who only plays or analyzes, the benefit is access to stronger, more efficient models without changing how you use the engine.
How to decide
If you contribute games, upgrade first; the training run's own timeline makes this a requirement rather than a preference. If you only download networks, check the networks page for the latest transformer release and compare it against your current model in a few positions you know well. For background on the engine itself, see KataGo Distributed Training and the project's GitHub releases.
How can I download and use the trained KataGo neural network models?
Download the trained networks from the run's Networks page, then point KataGo at the file with the -model flag (or the equivalent setting in your GUI). The models are the trained neural nets produced by the distributed self-play run; they are separate from the KataGo engine binary, which you get from the project's releases.
What you need
- Engine binary: a current KataGo release. The run's updates note that support for versions before v1.18.0 has been retired, and that older contributors were asked to upgrade before that cutoff. Use a recent release so your build matches the model format.
- Network file: the
.bin.gz(or uncompressed.bin) file from the Networks page. - A config: the standard
gtp_example.cfg/analysis_example.cfgshipped with the release, edited to set the model path and rules.
Basic usage
Command line, roughly:
katago gtp -model <network>.bin.gz -config gtp_example.cfg
Then connect your Go GUI (Sabaki, Lizzie, KaTrain, OGS review tools) to that GTP process. For batch analysis, run katago analysis with the analysis config instead.
Choosing a network
The run's updates describe a shift in model type, which matters for picking a file:
| Model type | Notes from the run's updates | Practical implication |
|---|---|---|
| Older convolutional nets (pre-transformer) | Trained through the earlier part of the run | Works with older engine versions; still usable, but no longer the strongest |
| Transformer nets (v1.17.0 and later support) | Described as stronger and more efficient than prior models; first ones uploaded to the run, with further fixes being experimented with | Requires a recent engine; the better default choice now |
Within a family, larger blocks/channels (the b40c768 style names) are heavier and slower per move but generally stronger; smaller ones are faster on modest hardware. Match the size to your GPU and time budget rather than always grabbing the biggest file.
Practical tips
- Check strength vs speed. On a laptop CPU, a mid-size net with more playouts usually beats a huge net with very few. On a decent GPU, the large nets pay off.
- Keep rules and board size in mind. The same net handles multiple board sizes and rule sets, so set the rules in your config to match the game you are reviewing; otherwise the evaluation is misleading.
- Update the engine when the run does. The updates show version cutoffs happening, so an old binary can silently stop being supported for contributing and may lag on new model formats.
- Contributing is optional. If you only want to analyze your own games, you do not need to join the distributed run; just download a network and run it locally.
For background on the engine itself, see KataGo Distributed Training and the project's GitHub releases for binaries and example configs. A sensible next step: download one current transformer network and one config, run a single game review, and compare its suggested moves against your previous setup before committing to a bigger model.
Hosting and Email
Cloudflare
Hosting
Google Workspace
Location
Pages, Search and Sharing
All bots
Registration details 
Identified technologies
User reviews (0)