Cryolite · Riichi (Mahjong Soul rules) · Open source

Kanachan

Not an engine you can play, but a kit for building one. Kanachan trains models on Mahjong Soul game records, with the stated goal of beating NAGA and Suphx.

GitHub ↗

Built by: Cryolite, personal project ¹
Rules: Riichi, Mahjong Soul standard rules ¹
Approach: Imitation learning on game records ¹
Strength: No trained model published ¹
Access: Code on GitHub, no licence file ¹²

Who built it

Kanachan is a personal project by GitHub user Cryolite, started in August 2021 ¹². Its stated aim is an AI that can beat the top existing engines such as NAGA and Suphx ¹.

Rules

Kanachan targets the rules of Mahjong Soul’s standard games, and its simulator is written to match them ¹.

How it works

Tiles are fed to the model as plain tokens through embeddings, with few hand-crafted features, and the author plans to scale up to more expressive models such as transformers ¹. Training starts with behavioural cloning (imitating moves in game records). It then uses what the README calls curriculum fine-tuning: the trained encoder is kept and only the decoder is swapped for a harder objective, such as score change or final placement ¹.

The author reported crawling about 65 million rounds from Mahjong Soul’s Gold Room and above as of August 2021 ¹. The project is split into C++ tools (log annotation, a simulator, a shanten calculator and wall reconstruction) and Python training code using PyTorch ¹.

Strength

No strength can be measured: the repository provides no training data and no trained models ¹.

How to access it

Clone the repository and bring your own data and compute. You must collect Mahjong Soul game records yourself; the README suggests a network sniffer as one way ¹. A Docker image of prerequisites and a dev container are provided for developers ¹.

The repository has no licence file ², so it grants no explicit right to reuse the code. Ask the author before building on it.

Sources

Facts checked against these sources on 24 October 2026.