Published October 3, 2026 · Updated October 16, 2026
Jungle Chess on Mistboard has a new top bot. It is KataGo-AnimalChess, hzyhhzy's KataGomo built on lightvector's KataGo, playing with the b10c384 network from Dandelion 4 by Kouza (lxsgx23). We play it here with hzyhhzy's permission, under the name KataGo.
KataGo learned Jungle Chess by playing itself, the way AlphaZero learned chess. A neural network proposes moves and judges positions, and a search checks its ideas. No other engine taught it.
Before putting it on the site, we played it against Misty, our own Jungle Chess bot, in an open challenge of 200 games at 1,000 visits a move. hzyhhzy's engine won 82, lost none and drew 118, a score of 0.705. Every win came from walking into Misty's den, and every draw was a repetition.
Mistboard's own rules code refereed every game, and each engine only proposed moves. The budgets were matched on time, about two seconds a move on one machine: Misty at its usual 5 million nodes on one thread, KataGo at 1,000 visits on four threads with the b10c384 network. Both engines stop at a fixed amount of search, so the games played on Modal are the games either would play here. KataGo's own repetition rule was switched off, so the referee's applied to both sides.
KataGo
Wins
Draws
Losses
Score
As red, 100 games
44
56
0
0.72
As black, 100 games
38
62
0
0.69
Total
82
118
0
0.705
The median game ran 127 plies. The 95% interval on the score is 0.67 to 0.74, about +125 to +180 Elo; the challenge asked for 0.55.
All 200 games are in one study, and the raw records are in one file, a line of JSON per game.
The tiger now jumps the river sideways
We found KataGo while searching GitHub for jieqi engines. It lives on the AnimalChess2025 branch of a repository named KataGomo, and nothing in that name says Jungle Chess, so an earlier search for Jungle engines had missed it.
Its tiger jumps the river sideways as well as lengthwise, the same leap as the lion. Ours jumped lengthwise only, a reading taken from English Wikipedia's article. Tencent's QQ游戏 rules, Leiden University's Dou Shou Qi page and KataGo's own source all give the tiger the lion's jump, and a match between an engine trained on one rule and an engine playing another would have measured the rule.
Since 21 September the tiger on Mistboard jumps both ways, in the rules engine, the bot, the analysis board, the rules page and the puzzles. One puzzle the new jump refutes was withdrawn. The change only adds moves, so every earlier game is still legal as played.
Two wins, read by both engines
We scored every position of two of those wins again with both engines, at the match settings. Each chart shows one number, KataGo's chance of winning, judged twice: once by KataGo and once by Misty. Where the lines part, the engines disagree about who is winning. A draw counts as half a win, and Misty's centipawns become a chance through the curve our analysis board uses, 1 / (1 + e^(−0.00368 × cp)).
Game 67: a wolf up, and lost
Misty, playing red, won a wolf on move 11 and read itself ahead for most of the game, while KataGo's score stayed at 50% and then climbed.
On move 29 Misty moved its cornered tiger from g8 to g9, and KataGo's score for itself went from 61% to 92% while Misty still read +111. KataGo's line for red starts with lion b1-b2 and keeps red at 37%. Misty only saw it on move 36, seven moves later.
Game 94: the slow squeeze
Misty, playing black, won a wolf on move 19 and read itself ahead for most of the next 30 moves, while KataGo's score crept from 50% to 63%.
On move 50 Misty moved its cat from c8 to c7, and KataGo went from 63% to 98%. Misty read +80 and found the forced loss two moves later. KataGo's line for black, elephant d6-d7, keeps black at 38%.
In these two games, KataGo saw the win seven moves and two moves before Misty did.
KataGo is the Go program, retrained by self-play
KataGo-AnimalChess is stock KataGo with the board and rules swapped for Jungle Chess and the self-play loop run as it was: 400 visits a move in self-play, a 10-block, 384-channel network trained by SGD on two GPUs, and rules randomised across the variants the Dandelion app supports, so one network plays all of them. The repository estimates under $50 of compute to reach top strength in a small game. The weights ship inside the app download, with no separate release and no stated licence.
Misty's evaluation is three hand-written terms: material, distance to the den and a trap penalty. A learned correction on top of it did not help.
Misty has no idea of a den race
Misty scores a position by material plus a bonus for each piece’s distance to the enemy den, so it learns who wins a den race only when the search reaches the den. In August we tried three ways to teach it, and none helped once the search budget was held fixed; races decide only 4.7% of games at full strength. Teaching our Jungle Chess engine the den race has the matches, and a game where depth found the race on its own.
Misty is still where you start
KataGo is at the top of the Jungle Chess bot list, above Misty. Misty stays the default and the easier opponent; pick KataGo when you want the stronger game. Game review and analysis for Jungle Chess still run on Misty.
On the site KataGo searches 150 visits a move, which takes about 1.5 to 3.5 seconds. At that setting it scored 19 wins, no losses and 31 draws in 50 games against Misty, 0.690, which cannot be told apart from the 1,000-visit result. The strength is in the network.
Thanks
KataGomo is open source at github.com/hzyhhzy/KataGomo. Thank you, hzyhhzy, for the engine and for letting us use it; lightvector, for KataGo; and Kouza, for the Dandelion 4 network.