Every capture is an explosion. The piece you take, the piece you took it with, and everything on the four points next to the capture are gone. Soldiers survive. Generals do not.

So there are two pieces you can never take: anything standing next to your own general. And the two advisors beside the enemy general are the two points the whole game turns on, because whoever captures on either of them wins on the spot.

What this is

Atomic Xiangqi is Chinese chess with atomic chess’s explosion. I adapted it to the 9 by 10 board in 2026 as a Mistboard original; as far as I can find, nobody had put the explosion on a xiangqi board before. It is live now on Mistboard, against eight engine levels or a friend, and it plays with an ordinary set from the ordinary array.

Atomic chess began on the German Internet Chess Server in 1995, from rules Klaus Knopper collected from friends who played it over the board. Lichess added it in 2015 and hosted 4.9 million games of it in 2021. On the chess board it is a game of sacrifices and early king hunts, with a first-move edge for White that is much bigger than in chess. On the xiangqi board it comes out differently, and the numbers for that are at the end.

The rules

A capture removes the capturing piece, the captured piece, and every piece on the four points orthogonally next to the capture point. Diagonal neighbours are not touched.

Red’s chariot takes the horse on e6. The chariot, the horse, and the chariot and cannon on the two points beside it all go.

Soldiers survive a blast. A soldier leaves the board only by being captured itself, though a soldier that captures explodes like anything else.

Red’s chariot takes the horse on e5. The cannon on f5 goes; the soldiers on d5 and e6 are on the lines too and stay. A soldier that captures explodes like anything else.

A general does not survive a blast. Taking a piece beside the enemy general wins on the spot, and taking a piece beside your own general is not a legal move, so a general never captures. At the start each general has two advisors beside it, and those are the points the game turns on.

The chariot on d9 may take the advisor, and that ends the game: the blast reaches the general on e10. The chariot on f2 may not take the horse, because the blast would reach Red’s own general on e1.

A cannon’s capture is the one exception: it removes only the cannon and its target. Nothing beside the target is touched, and neither is the piece the cannon jumped.

A cannon’s capture is the one that does not explode. The cannon on e2 jumps the soldier on e4 and takes the horse on e7; the chariot on d7 and the cannon on f7 beside the target are untouched, and so is the soldier it jumped.

Everything else is xiangqi, with one word widened: you are in check whenever the opponent could remove your general next move by any means, so a chariot bearing on the advisor beside your general is giving check, and the board says so. That matters for xiangqi’s perpetual-check law, which stands. The rules page has all of this in one place, with a sample game.

Why the cannon is the exception

The first version had the cannon explode like everything else, and engine play found two things wrong with it. The first was power. From the starting array each cannon can jump the enemy cannon, take the horse behind it and blow up the chariot and elephant beside that, three pieces on move one, and every best-play game opened with both sides doing exactly that.

How every engine game opened when the cannon exploded like everything else. Each cannon jumps the enemy cannon, lands on the horse and explodes it with the chariot and elephant beside it; the cannon dies too. Three pieces for one on each wing, before either side has made a plan.

The second was a drawing resource. A chariot on the file of an advisor threatens to blow up the general; the only parry is a cannon blocking the file; and with cannons exploding, that block is also a shot back over the chariot at the attacker’s own advisor, so the chariot has to hop to the other advisor file, the cannon follows, and the position repeats. The defender can never decline, and in the engine games a side that had lost a chariot could hold the draw this way against a far stronger opponent.

The three-move rush and why it fails. Red’s chariot on d2 is about to take the black advisor on d10 and explode the general. But a cannon captures by jumping exactly one piece, and the chariot on d2 is the one piece between the cannon on d8 and Red’s own advisor on d1. Black shoots first: advisor, chariot and the general on e1 all go.

Removing both took two rules, and neither works alone. The cannon’s capture does not explode, which ends the opening massacre and leaves the block standing but harmless. And a threat to blow up the general counts as check, which makes the hop that remains a one-sided perpetual check, and under xiangqi’s own law the side giving it loses. With both in place the engine plays on at minus a third of a pawn rather than take the draw, and the cannon comes out with a job it did not have: it blocks a chariot without threatening it, and the chariot side has to find a plan.

A sample game

Games are shorter and more violent than xiangqi’s. Of ten decisive engine games between a strong engine and a weaker one, six ended with a general blown up and four with checkmate, the shortest in 23 moves and the longest in 73. Openings still look like xiangqi, because the first explosion is expensive for whoever starts it; the engine opens Cb5, Ri3, Ra3 or Hg3, and the first blast comes in the middlegame.

Here is the shortest game: Fairy-Stockfish at ten million nodes as Black against itself at one million. Red’s general steps into the corner of the palace on move 6, a horse takes a horse and the cannon beside it on move 8, and the last move is a chariot taking the advisor beside the general, which is how most games here end.

The soldier is worth a second game. It survives every blast, so it is the one piece that can stand next to an explosion and stay, and it is also the cheapest thing to capture with, since whatever captures is lost anyway. Here is a soldier about to take a chariot on c9. Step forward once: the chariot, the elephant on c10 and the horse on c8 all go, for a soldier.

What the engine games say, and what they cannot

Every game was refereed by a rule kernel and played by a patched Fairy-Stockfish, and the two have to agree on every move count before a number is read. Under these rules the stronger engine wins 29 of 32 games across two strength gaps without losing one; of the 44 first moves from the array, none wins for Red at a million nodes a side and three lose; and none of the 19 repetition draws on file is a one-sided check. Best play is still a draw, which is the normal condition of a board game. The difference from the first version is that the draws have no forced line in them. The variant lab has the kernel, the patch and the commands behind each number.

What the engines cannot say is what the game is like between people. Whether equal players grind (four of eight games between equal engines at a low node count were draws), whether a xiangqi player takes to a cannon that trades and blocks instead of clearing a rank, and whether “a chariot on your advisor file is check” is a rule you learn by losing to it once: only people can answer those. If you find a forced line, send it and I’ll add it here with credit.

Play Atomic Xiangqi The rules The twelve engine games