Life In 19x19

@lovego
The latest release at https://github.com/aki65/aki65.github.i ... g/v1.20.13 is built with Qualcomm's latest NPU drivers (as of 2025-04-30), so you could have a try with the Snapdragon 8s gen 4 (although it's still not mentioned in the release notes)

According to Qualcomm's release notes the new chip isn't supported yet. I expect, that it will be added soon.

I'm working on a slightly improved weight optimization. When it's ready, I will publish new weights.

I'm also implementing a new feature, but it's difficult to explain without seeing it, so I will postpone any description until it's ready.

Katago 1.16.0 is on my todo list, but with low priority ...

The latest version https://github.com/aki65/aki65.github.i ... ag/v1.20.9 works on 8 Elite chip.

I uploaded the network here: https://github.com/aki65/aki65.github.i ... ag/v1.20.5

I made a new version available at https://github.com/aki65/aki65.github.i ... ag/v1.20.1 , which includes support for KataGo's human SL network.

Sorry for the late response, I have been on vacation.

I'm currently working on integration of KataGo's human SL network into BadukAI.

Adding sayuri or pachi would be an interesting programming task, but it's a lot of work without much benefit for users, so it does not have high priority for me.

I don't have a SD 8s gen3 phone to test it, so I can't give any guarantees. But I am quite optimistic, since (according to Qualcomm's documentation) the DSP of the SD 8s gen 3 has same architecture as SD 8 gen 2.

Regarding the part of KataGo's functionality that BadukAI uses, there have been no significant changes from 1.14 to 1.15.1, so it's not urgent to upgrade. But of course I will do so sooner or later ...

I made a new version (1.20) available on play store https://play.google.com/store/apps/details?id=net.kir.baduk_ai and github https://github.com/aki65/aki65.github.io/releases/tag/v1.20.0 which wraps up all the changes in the pre-releases on github since last September:

customization of rules ...

I put the optimized weights for that network here: https://github.com/aki65/aki65.github.i ... g/v1.19.18

@lovego
The optimized version of the new 28b network is here: https://github.com/aki65/aki65.github.i ... g/v1.19.16
How many visits/s does it yield on your device (in the reference position of my last network post) ?

The optimized 18b and 28b networks can only be loaded on Qualcomm processors. Huawei chipsets do not support all operations on NPU that are needed by these networks.

You can, however, load original KataGo networks ( either the built-in 10b or any .bin.gz file from https://katagotraining.org ...

The optimized weights are here: https://github.com/aki65/aki65.github.io/releases/tag/v1.19.13

If you want to let them play against each other, keep in mind that on android the optimal thread count is 2 (for both 18b and 28b)

By the way, it would be interesting to know the speed of the 28b ...

I optimized kata1-b18c384nbt-s8040575488-d3801933292 just a few weeks ago (available at https://github.com/aki65/aki65.github.i ... ag/v1.19.6) and it's only 10 ELO weaker than kata1-b18c384nbt-s8526915840-d3929217702, so that one should be sufficient.