Gomocup 2025, the 26th tournament (June 6th-8th, 2025)

The Gomocup 2025 took place on June the 6th-8th, sponsored by Tech Center of Lightspeed & Quantum Studios Group, Tencent.

There were four cloud servers whose configurations were the same:

  • Tencent Cloud Servers SA2.8XLARGE64 (Windows Server 2025 Datacenter, x64 32 vCPU (AMD EPYC 7K62, 2.6/3.3GHz), 64GB RAM)

The openings for Gomocup 2025 were chosen by the following people (sorted alphabetically according to last names):

  • Alexander Bogatirev - Gomoku player, member of Gomoku Committee RIF, member of organizers committee of Russian Gomoku Championship, winner of Russian Gomoku Cup 2016.
  • Nguyễn Minh - Gomoku player and opening analyzer, who experienced in Gomoku swap2.
  • Zijun Shu - Gomoku player and AI researcher, contributor of several Gomoku AIs.
  • Tao Tao - Renju theory researcher, who published some researches and new designs of renju opennings, participated in promotion of Renju and translated several Japanese renju books.
  • Qichao Wang - Renju expert, 4 dan, who has done a lot of researches in renju openings, and won the World Correspondence Renju Championship 2019.

Thank you all!

Same as last year, we prepared 12 openings for every rule.

  • Openings for the freestyle (20x20) league:

  • Openings for the freestyle (15x15) league:

  • Openings for the standard league:

  • Openings for the renju league:

  • Openings for the caro league:

What is new?

  • Updates
    • AlphaGomoku (Kozarzewski) 2025 - New neural network architecture.
    • Ego-Zero 2025 - This version is based on MCTS and ResNet.
    • Embryo 2025 - Better support for AVX2 instruction set. Some speedup in search.
    • Jax 2025 - The Monte Carlo tree search plus neural network model solution from last year is used, but the models are more powerful and there are fewer of them, and the four rules (Freestyle, Standard, Renju, Standard Caro) with a board size of 15 use the same model; one engine supports all groups of Gomocup2025 competitions; and a memory problem in multi-threaded search is fixed.
    • Rapfi 2025 - New NNUE architecture featuring a reduced weight size and lower memory footprint. Added support for AVX-VNNI, AVX-512, and AVX-512 VNNI instruction sets.
  • New AIs
    • Figrid - developed by Bobo Wu from China. First version, really weak.
    • TPTC - developed by Bruno Yerro from France. TPTC is written in python.

Starting this year, we increase the memory limit to 1GB and the AI size limitation to 256MB. Note that every submission we received in or before 2024 does not exceed 256MB.

There were 2 freestyle (20x20) groups (i.e., Freestyle-20 1 and 2), 2 freestyle (15x15) groups (i.e., Freestyle-15 1 and 2), 1 fastgame group (Fastgame), 2 standard groups (i.e., Standard 1 and 2), 1 renju group (Renju), and 1 caro group (Caro) in Gomocup 2025. Elo rating system is used to evaluate AIs' strength. Time per move/per match was 5s/120s for fastgame, 300s/1000s for final leagues (i.e., Freestyle-20 1, Freestyle-15 1, Standard 1, Renju, and Caro), and 30s/180s for the rest (i.e., Freestyle-20 2, Freestyle-15 2, and Standard 2).

AIs were divided into different freestyle and standard groups according to the placement in the last tournament. For Freestyle-20 2, Freestyle-15 2, and Standard 2, the top 4 AIs will move up to the next group (i.e., Freestyle-20 1, Freestyle-15 1, and Standard 1, respectively). If the top k (k>4) places were all taken by new or updated AIs in a group, then all these k AIs would advance to the next group.

In Fastgame, Freestyle-20 2, Freestyle-15 2, and Standard 2, we do not match every pair of AIs. Instead, for old AIs, we only match every pair of them whose last year's Elo difference was no more than 400. We still match each brand new/updated AI with all AIs.

The unlimited tournament, which has no hardware constraints, will not be held this year, as we only received one submission for this league.

There were technical difficulties like every year.

  • The server setup took a bit longer this year. In particular, Windows Server 2025 requires additional configuration to enable support for .NET Framework 3.5, which is needed by some AIs. We also noticed that the disks on the competition servers were nearly full during the competition, so we added additional storage in time to avoid potential risks.

TITO, DEEPFIRE, HEWER, and WHOSE were the top 4 in the freestyle (20x20) 2 league and advanced to freestyle (20x20) 1.

The latest RAPFI retained championship title in the freestyle (20x20) league this year. The new ALPHAGOMOKU (Kozarzewski) was the second, and the updated JAX was the third.

WHOSE, TITO, DEEPFIRE, and ARBILLO were the top 4 in the freestyle (15x15) 2 league and advanced to freestyle (15x15) 1.

Same as last year, the winner of the freestyle (15x15) league was RAPFI. The second was JAX, and the third was ALPHAGOMOKU (Kozarzewski).

The winner of the fastgame league was RAPFI. The second place was taken by ALPHAGOMOKU (Kozarzewski), and the third by JAX.

WHOSE, TITO, SLOWRENJU, and HEWER were the top 4 in the standard 2 league and advanced to standard 1.

RAPFI won the standard league. The second place was taken by ALPHAGOMOKU (Kozarzewski), and the third by JAX.

The winner of the renju league was RAPFI. ALPHAGOMOKU (Kozarzewski) took the second place. The third was JAX.

The winner of the caro league was ALPHAGOMOKU (Kozarzewski). JAX took the second place. The updated EMBRYO took the third.

You can download complete results and openings here.