Mahjong Engines
The index of mahjong AI

Every mahjong engine, read side by side.

Who built it, which rules it plays, how it thinks, and how strong it really is. From Tenhou 10-dan research bots to the open-source models you can run tonight.

7Engines indexed
6Rule sets tracked
4Protocols & envs

What an engine sees

Riichi · East 2 · Turn 9
Selected discard
White Dragon
Isolated honor. Leaves 2–5 sou two-sided wait.
Policy weight
91%

Illustrative policy output, not taken from a specific engine. Tap a tile to see its weight.

Coverage by rule set

Engine research is lopsided. Riichi has a deep bench; most other tables have barely been touched.

Japanese Riichi

7 engines · 2 envs

Chinese Official

Botzone ladder

Hong Kong

Not yet catalogued

Taiwanese 16-tile

Not yet catalogued

American (NMJL)

Not yet catalogued

Indian

Not yet catalogued

The engines

Strength figures are as reported by each project. Click through for papers, repos and match logs.

Engine Rules How it thinks Strength Access

Suphx

Microsoft Research Asia
RulesRiichi
How it thinksDeep reinforcement learning with oracle guiding and global reward prediction.
StrengthTenhou 10-dan
AccessPaper only

LuckyJ

Tencent AI Lab
RulesRiichi
How it thinksReinforcement learning combined with game-theoretic search.
StrengthTenhou 10-dan
AccessResearch

NAGA

Dwango Media Village
RulesRiichi
How it thinksDeep neural networks trained on top-level human games.
StrengthTenhou 10-dan
AccessPaid review

Mortal

Equim · open-source
RulesRiichi
How it thinksDeep RL model in Rust and PyTorch. Speaks mjai; powers community log reviewers.
StrengthCommunity ladders
AccessAGPL-3.0

Kanachan

Cryolite · open-source
RulesRiichi
How it thinksTransformer trained on large sets of Mahjong Soul game records.
StrengthSelf-reported
AccessOpen source

Akochan

critter · open-source
RulesRiichi
How it thinksHand-built expected-value search over tile efficiency and danger.
StrengthCommunity ladders
AccessOpen source

Bakuuchi

University of Tokyo
RulesRiichi
How it thinksMonte Carlo simulation guided by learned opponent models.
StrengthTenhou Tokujou
AccessResearch

Protocols & environments

The plumbing engines plug into: message formats for live play, simulators for training, and ladders for ranking.

Protocol · Riichi

mjai

JSON-over-socket protocol for riichi bots. Most open-source engines and log reviewers speak it.

{"type":"dahai","actor":0,"pai":"P","tsumogiri":false}
Environment · Riichi

Mjx

Fast riichi simulator built for reinforcement learning, with Tenhou-compatible rules and mjai interop.

Platform · Chinese Official (MCR)

Botzone

Peking University's bot arena. Hosts Chinese Standard Mahjong matches and past IJCAI competitions.

Toolkit · Simplified

RLCard

Card and tile game toolkit with a simplified mahjong environment. A common starting point for coursework.

How strength is measured

Mahjong is noisy. A hundred hanchan prove little, so we note the ladder, the sample size and who ran it.

Human ladder

Tenhou dan

Rank earned in Tenhou's Tokujou or Houou room. 10-dan is the ceiling only a handful of humans and bots have reached.

Stable rating

Stable dan & R

Estimated long-run rank from placement rates. Better than peak dan because it washes out lucky runs.

Head to head

Duplicate matches

Engines replay the same walls from every seat. Cuts variance enough to separate close bots in thousands of games, not millions.

Built an engine?

Hong Kong, Taiwanese, American and Indian tables are wide open. Send us the repo, the rule set and a match log, and we'll write it up.

Submit an engine