Here are some fun extra networks with special training that are not part of KataGo's main training run.
Here are some fun extra networks with special training that are not part of KataGo's main training run.
Human SL Network (July 2024)
[Network file] [Raw pytorch checkpoint]
This is a supervised learning net that was trained on a large number of human games to predict human moves across players of all different ranks and time periods! It was originally released with v1.15.0, and that release page has some pretty pictures and further information about its training. The HumanSL model can be used in various experimental ways to try to make KataGo play or analyze games in a more human-like way with GTP, see comments and example config here, and/or the analysis engine command, see documentation here.
For developers, it may be possible to come up with other creative custom uses. Such as using the model for player rank estimation, or as a filter in teaching tools for mistakes (among two otherwise equal "mistakes", one might be much less worth reviewing if the HumanSL model thinks that e.g. even a moderately stronger player would be likely to miss the move than one which the model predicts a high chance that players of the given rank should get it right). The model can also be configured to predict pro games based on historical year, as well as how players would play in handicap games.
"FD3" Network (trained late 2024, posted April 2025)
[Network file]
This is a network privately finetuned and used by a number of competitive KataGo users originally in 2024 with learning rate drops to much lower than the higher learning rate maintained by the official KataGo nets, and which has since been released for public download. This network is probably similar in strength or slightly stronger than the official networks in normal games as of April 2025! Although it might not stay as up to date on certain blind spot or particular misevaluation fixes as various such training continues ongoingly through 2025.
Lionffen b6c64 Network (posted April 2025)
[Network file]
Trained by "lionffen", this is a heavily optimized very small 6-block network that in normal games may be competitive with or stronger than many of KataGo's historical 10-block nets on equal visits, while running much faster due to its tiny size! It has been trained specifically for 19x19 and might NOT perform well on any other board sizes. Additionally, due to being a very shallow net (only 6 residual blocks), it will have too few layers to be capable of "perceiving" the the whole board at once, so like any small net, it may be uncharacteristically weak relative to its strength otherwise in situations involving very large dragons or capturing races, more than neural nets in Go already are in such cases.
Finetuned 9x9 Net (October 2023)
[Network file] [Raw pytorch checkpoint]
This net is likely one of the strongest KataGo nets for 9x9, even compared to nets more recent than it! It was specially finetuned for a few months on a couple of GPUs exclusively on a diverse set of 9x9 board positions, including large trees of positions that KataGo's main nets had significant misevaluations on. This was also the net used to generate the 9x9 book at https://katagobooks.org/.
Do not expect this net to be any good for sizes other than 9x9. Due to the 9x9-exclusive finetuning, it will have forgotten how to evaluate other sizes accurately.
If you're interested, see the original github release post of this net for more training details!
Short Distributed Test Run "Rect15" Final Net (December 2020)
[Network file] [Raw tensorflow checkpoint]
Just for fun, this is the final net of a short test run for KataGo's distributed training infrastructure, before the official run launched. It was trained on a wide variety of rectangular board sizes up to 15x15, including a lot of heavily non-square sizes, such as 6x15. It is only a 20 block net, and was trained for far less time than KataGo's main nets. It has never seen a 19x19 board, and will be weak on 19x19 by bot standards, but may still be very strong by human amateur standards and still play reasonably by sheer extrapolation.
Strong Large Board Net "M2", b28c512nbt size (May 2025) (Trained by Friday9i)
[Network file] [Raw pytorch checkpoint]
This is a strong net finetuned by "Friday9i" for months starting from KataGo's official nets to be vastly stronger on boards larger than 19x19! It should be stronger than the official nets by many hundreds of Elo for board lengths in the high 20s, and virtually always winning on board lengths in the 30s, where the official nets start to behave nonsensically. As of mid 2025, this net is the ideal net to use for large board play for the "+bs50" executables offered at KataGo's latest release page that support sizes up to 50x50.
According to Friday9i, even this net might not be 100% reliable on score maximization or finishing up dame or other small details for board lengths in the high 30s or in the 40s but should still behave overall reasonably and play fine. See this forum post for more stats and details. Enjoy!
Kill-all Go Net (October 2024) (Trained by hzyhhzy)
Link: https://github.com/hzyhhzy/KataGomo_fork/releases/tag/LifeGo_20241025
Developed by "hzyhhzy", this is a finetuned/transfer-learned KataGo net along with specially modified KataGo source code that is trained for some positions where Black starts with either 17-20 stones, or alternatively the entire first line of the board, where Black is trying to keep the entire board while White tries to make a living group anywhere.
According to their release notes, although it is designed for alternative code, the network should work with the regular KataGo executable as well if komi is set to 20 (rather than any value near 360), since that's where the input feature encoding of the normal KataGo aligns with the modified encoding that this special net would expect to receive from the hzyhhzy's modified KataGo, although the scoring in variations with passes will be messed up, so it may be preferable to just use the modified code directly.
Strong Igo Hatsuyoron 120 Net, 40 blocks (February 2022) (Trained by Friday9i)
[Network file]
Trained by "Friday9i" for close to a year, this is one of the strongest publicly-available networks specially trained to understand what has been termed the most difficult problem in the world, problem number 120 in a classic problem collection from Inoue Dosetsu Inseki dating back to the 1700s.
This network, its earlier versions, and/or some further never-publicly-released networks along with extensive human work and analysis by Thomas Redecker and other researchers are responsible for the significant discoveries and refinements in human understanding of the problem in the years after initial new moves were discovered by KataGo in 2019. The effort to analyze and solve this problem has been an amazing effort across the years, and is documented in detail by Thomas Redecker at https://igohatsuyoron120.de/.
First Igo Hatsuyoron 120 Net, 20 blocks (December 2019)
[Network file]
This is the very first network that was specially trained to understand what has been termed the most difficult problem in the world, problem number 120 in a classic problem collection from Inoue Dosetsu Inseki dating back to the 1700s. Although trained far less than later networks like the 40-block network above contributed by later independent developers, this first network was able to discover a few key new moves and variations, which was the start of major new advances in the human understanding of the problem in 2019, explained extensively at https://igohatsuyoron120.de/ and also originally featured at https://blog.janestreet.com/deep-learning-the-hardest-go-problem-in-the-world/.