Suphx
The first AI to reach 10 dan on Tenhou. Suphx (short for “Super Phoenix”) paired supervised learning on top human games with self-play reinforcement learning, and published the method in full.
Paper ↗- Built by
- Microsoft Research Asia 1
- Rules
- Four-player Japanese riichi 1
- Approach
- Deep CNNs, supervised then self-play RL 1
- Strength
- Tenhou 10 dan, stable rank 8.74 1
- Access
- Paper only
Who built it
Suphx was built at Microsoft Research Asia. The paper lists ten authors: Junjie Li, Sotetsu Koyamada, Qiwei Ye, Guoqing Liu, Chao Wang, Ruihan Yang, Li Zhao, Tao Qin, Tie-Yan Liu and Hsiao-Wuen Hon, with four of them working as interns at the time 1.
Microsoft announced the result on 29 August 2019 and credited the team led by Hsiao-Wuen Hon, then head of its Asia-Pacific R&D Group 2. The name is short for Super Phoenix 12.
Rules
Suphx plays four-player Japanese (riichi) mahjong, the rule set used on Tenhou 1.
How it works
Suphx uses deep convolutional neural networks. It has five models: one each for discards, riichi, chow, pong and kong 1. The networks are first trained by supervised learning on logs of top human players, then improved by self-play reinforcement learning with a policy gradient algorithm 1.
The paper adds three techniques 1:
- Global reward prediction. A predictor estimates the final game reward from the current and previous rounds, so the agent learns from the game result rather than from a single round’s score.
- Oracle guiding. Training starts with an oracle agent that can see opponents’ hands and the wall. That perfect information is gradually removed until the agent sees only what a normal player sees.
- Run-time policy adaptation (pMCPA). The trained policy is adjusted to the specific round as it unfolds, using simulations.
The version that played on Tenhou was trained on about 2.5 million games and ran without run-time adaptation, because Tenhou’s time limit per action made it too slow 1.
Strength
On Tenhou, AIs may only play in the expert room; the phoenix room is reserved for human players of 7 dan and above 1. Suphx played over 5,000 games there, reached 10 dan and recorded a stable rank of 8.74 1. The paper calls it the first and only AI on Tenhou to reach 10 dan, and rates it above 99.99% of ranked human players 1. Microsoft noted that only about 180 people had ever reached 10 dan 2.
| Player | Games | Record rank | Stable rank |
|---|---|---|---|
| Bakuuchi | 30,516 | 9 dan | 6.59 |
| NAGA | 9,649 | 8 dan | 6.64 |
| Top humans (10 dan, pooled) | 8,031 | 10 dan | 7.46 |
| Suphx | 5,760 | 10 dan | 8.74 |
By stable rank, Suphx was about 2 dan stronger than Bakuuchi and NAGA, the two strongest AIs before it 1. Tencent’s LuckyJ later reported a higher stable rank.
How to access it
Microsoft published the method as a paper. We found no released code, model weights or public way to play against Suphx; Microsoft’s announcement frames it as research whose techniques could later be applied to real-world problems 2.
Sources
- Li et al., “Suphx: Mastering Mahjong with Deep Reinforcement Learning”, arXiv:2003.13590 (2020)
- Microsoft News, “Mastering Mahjong with AI and machine learning” (29 August 2019)
Facts checked against these sources on 24 October 2026.