Recent posts

The 2022 women's world champion and a double gold medallist at the 2023 Asian Games. Her title run in Kuching and her 2026 season, five games on the board, 32 games analysed.

On 11 October we stopped new games of Fortress Xiangqi and Atomic Xiangqi so we can redesign them. In Fortress the drops took over a small board; Atomic only stopped looping after the cannon got its own exception. Finished games still replay.

In banqi a chased piece often has one safe square, so a chase can go back and forth until the position repeats, and under plain threefold repetition the chaser alone decides the draw. Tournament rules allow it and apps cap the chase at 7 or 10 moves. Mistboard forbids the chase move that would repeat a position a third time, and leaves chases that go somewhere alone.

We searched 2,939 positions from 50 jieqi games for puzzles. 58 had one clearly best move and 15 of those turned on a hidden piece. None of the 15 passed the rules our xiangqi puzzles go through, and we stopped looking.

A horse and a soldier beat a general with both advisors and both elephants, and the slowest win in that endgame takes 65 moves. Seven records from xiangqi endgame tablebases, with the lines, and five of them end with the loser having no legal move.

We put Benedict Chess’s rule on the xiangqi board: nothing is captured, and every enemy piece your moved piece attacks joins you. The game has no endgame and no draws, and the first mover wins 78% of engine games, nearly all of it from one cannon move. The pie rule looks like the repair.

Xiangqi basic endgames, which win and which draw: a table of 35 common endings with the result and the key idea for each, every result checked against the chessdb.cn tablebase, and every position ready to practice against the computer.

A one-character bug in Mistboard's xiangqi repetition key let a game end in a threefold draw after a position had occurred only twice.

Red’s first Duck Xiangqi turn reaches more positions than a full xiangqi move by each side. About half of them leave the opponent exactly the same choices, so a search can skip them. What is left still cuts how far an engine sees almost in half.

We ran an exhaustive solver over every short continuous-check composition in 適情雅趣 (1570), 渊深海阔 (1808) and a modern collection: 172 problems, asking whether the printed solution is the only one. Forty-five are. Seventy have a second winning continuation somewhere in the tree, fourteen have a second first move, nine can be mated faster, and three records turned out to be corrupt: the position as recorded cannot be the one the composer meant.

A Jungle Chess engine that learned by playing itself now sits above Misty. In 200 games against Misty it won 82, lost none and drew 118.

Our Pikafish jieqi bot rated some reveals far better than they were. Two bugs caused it. Both are fixed, at little or no cost in strength.

Hong Kong's first Asian individual champion and its 2022 world championship finalist. His last fifteen months and the 2022 run, six games on the board, 50 games analysed.

An open-source jieqi engine with its own neural network now sits above Pikafish Level 8. It beat full-strength Pikafish 248 to 136 in 400 games.

The first world champion from outside China, unbeaten in Shanghai and winner of the final with black. His title run and the year since, five wins on the board, 38 games analysed.

Since August 23, people have played our jieqi bot 596 times and beaten it 14 times. Here is how those games were won, and seven new levels for everyone else.

Since June 1, people have finished 1,000 games on Mistboard, most of them in September. Thank you for playing.

Beat the world champion 12 to 8 in a ten-game match in January, then the rest of his season through the first stage of China's top league. Six games on the board, all 75 analysed.

MistyBanqi played itself 200 times at ten million nodes a move under the Taiwanese competition rules. On the engine’s own point scale, a lead under ten points is a coin flip at every stage of the game, and the side that loses its general first loses 87% of the time.

Play Pikafish, the open-source xiangqi engine built from Stockfish, in your browser. Free, no account, no download. Choose a level, play it at jieqi, and review your games with it.

In 144 of 4,121 common opening positions, the move club players choose most often loses winning chances against the engine at twenty million nodes. Ten of them, with what the masters played in each era, and a study holding all 144.

31 games without a loss across the league qualifier, the Shanghai Cup and the first league stage, August to September 2026. How he plays, five games on the board, all 31 analysed.

We put Horde on the xiangqi board and measured it before building anything: twelve start arrays, two soldier rules, each played four times by an engine against itself. With xiangqi’s own soldier the army wins every game by one trick; give the soldier the crossed move from the start and the game becomes a siege that draws three times in four. We are publishing the measurement, not the variant.

We put antichess on the xiangqi board and measured it before designing anything. Black has two moves to find in the opening, Red cannot go wrong, and then the palace keeps five pieces a side out of reach. Every other opening loses, 63 of them provably. We are publishing the measurement, not the variant.

My fog of war chess engine recorded the en passant square under standard chess legality, in a variant with no such rule. En passant availability decides what a pawn can see, so the true board fell out of the engine's belief and it finished the game against 41 worlds that did not exist.

Chinese chess with one duck both players share. The rules take a minute to read. What decides your first game takes longer, starting with the fact that the duck you place is a cannon platform for your opponent and never for you.

A solver found a real mate and was told to try again. 382 served puzzles could do that. How the miner admitted them, the patch that half-fixed it, the rule lichess uses instead, and what came out of the corpus.

Jieqi has no opening book. It has an argument about the first move, running on Chinese forums among players with thousands of games, never written down in English. Why a face-down piece is a one-shot option you can waste, five openings ranked, and the pawn push weighed against the crossed cannon on all six reveals.

Play jieqi against the engine or a friend, free and without an account, then review the game with analysis that separates your choices from your luck.

A miner that reads real xiangqi games, finds the moves people got wrong, and keeps the positions where exactly one move wins. About one blunder in nine survives it. Here is the algorithm, the code, and some of what it kept and threw away.

Between 2024 and 2026 the Chinese Xiangqi Association sanctioned 49 people for buying and selling games, and a Hangzhou court convicted six grandmasters of bribery. What happened, why it paid, who ruled, and what is still unproven.

Every winner of the Xiangqi World Championship since 1990, why the Chinese national title is the harder one, and how a Vietnamese player took it out of China for the first time in 2025.

Every winner of the Chinese national xiangqi championship since 1956, and an annotated game for thirteen of them. Plus the nine hundred years before the title existed, and the decade that has been struck from the record.

Verified titled players get a gold badge beside their name, a coaching page students can find, and a front page that will carry their work. Verification takes about two minutes.

Red’s opening cannon reaches the riverbank first, one move from firing down any of five files, and in fog you never see it coming. Whether that breaks the game came down to one elephant move, one poisoned defense, and a coin flip we priced with the engine.

Half the moves in banqi, jieqi, and flip jungle are dice rolls, so a chess-style review blames you for variance. Mistboard’s game review splits every flip into the decision and the tile: luck-stripped accuracy, a luck line on the advantage graph, and what 52 human-versus-engine games say about who really earned their wins.

Flip Jungle is 4×4 Jungle Chess with every animal face-down. A near-perfect engine shows how much skill survives the flips: 512 times the search buys about 242 Elo and still loses 23% of its games.

Misty is Mistboard's Fog of War chess engine: how it sees, searches possible boards, avoids hidden catastrophes, and where the current version stands.

MistyBanqi is the engine you play in banqi on Mistboard: a classical search engine with a hand-written evaluation. How it thinks, and the blind spot worth knowing: it can draw a game it has already won.

How Mistboard keeps hidden information on the server: canonical state, seat-scoped views, private live rooms, and public postgame review.

Stockfish 18 rankings for all 960 Chess960/Fischer Random starting positions. Most balanced: SP-418 and SP-563; standard chess is near median.