Home › Apps › Chess Ta! › Chess variants
Chess variants in Chess Ta!
Seventeen ways to play, all offline, all on an engine written from scratch in C.
17 variants Offline One-time purchase
Each variant below is a separate set of rules with its own move generator, not a cosmetic mode — the engine really does play Atomic differently from Antichess. Sixteen of the seventeen have an AI opponent; Bughouse is pass-and-play only, because it needs four seats and two boards at once. Most can also start from a shuffled Chess960 back rank.
Classic variants
King of the Hill
The four centre squares win the game. March your king to d4, d5, e4 or e5 and you win on the spot, with checkmate still available the usual way. It turns the king from something you hide into something you run, and the centre stops being a place you fight over and becomes the finish line.
Three-Check
Check the enemy king three times and you win, no mate required. Checkmate still ends the game normally, so there are two ways to finish. Sacrifices that would be reckless in ordinary chess become sound if they buy a check, and a king that has been checked twice can never relax again.
Racing Kings
Nobody may give check, in either direction, and the first king to reach the eighth rank wins. Every piece starts off the back ranks and exists only to clear a path or block one. It plays less like chess than like a race where both runners can trip each other, and a single tempo usually decides it.
Horde
Thirty-six white pawns against one full black army. White has no king at all and wins by mating Black's; Black wins by capturing every last pawn. The pawn mass is slow but unstoppable if it is left alone, so Black spends the whole game deciding which part of the wall to break before it arrives.
Antichess
You win by losing everything. Captures are compulsory whenever one is available, and the king is an ordinary piece with no special status — it can be taken, and a pawn may even promote to one. Forcing your opponent to capture is the whole game, and a material lead is a problem, not an advantage.
Atomic
Every capture detonates, wiping out the capturing piece, the captured piece and the pieces standing around them. Blowing up the enemy king wins outright, so mate is not required — but the same blast can take your own king with it, which makes some perfectly legal captures fatal to the player making them.
Crazyhouse
Nothing you capture is ever really gone. Captured pieces switch sides and can be dropped back onto any empty square as a move, so material returns to the board in your opponent's hands. Attacks never quite die out, defence needs actual squares rather than just extra pieces, and the board only gets busier.
Duck Chess
Move a piece, then move the duck. The duck sits on one square and blocks it for everyone, and it moves every single turn. There is no check in this variant — the king can be captured like anything else — so threats have to be seen rather than announced, and the duck can shut down the only escape square.
Dark Chess
You see only the squares your own pieces can move to. The rest of the board is unknown, including where the enemy king is, and you win by capturing it rather than by mate. Also known as Fog of War, though we use its original name. Every move is a guess informed by what you last saw, and by what has stopped being visible.
Bughouse
Two boards, four players, two teams. Every piece you capture is handed to your partner, who can drop it onto their own board. You end up playing your own game and negotiating theirs at the same time. This is the one variant with no AI opponent — it needs four seats, so it is pass-and-play only.
4-Player Chess
Four armies on a cross-shaped board, playing as two teams: Red and Yellow against Blue and Green. Turns pass around the board, so the position in front of you changes three times before you move again, and a threat you ignore has two other players' worth of time to become fatal.
4-Player Free-for-All
The same cross-shaped board and four armies, but nobody is on your side. A mated player freezes in place and their pieces stay on the board as obstacles, while everyone still standing plays on. Attacking is expensive here, because the two players who are not involved get to profit from whatever you spend.
New in version 1.4.0
Five rule-tweak variants, each changing exactly one thing about ordinary chess. All five also play Chess960.
No Castling
Chess with castling deleted, an idea Vladimir Kramnik put forward to get more variety out of the standard opening. Everything else is untouched. The king has to walk to safety, which takes moves nobody can spare, and centuries of opening theory built around a quick castle stop applying immediately.
Torpedo
Any pawn may advance two squares from any rank, not only from its starting square. Pawns arrive roughly twice as fast, passed pawns turn into genuine emergencies, and a structure that looks safely blocked can be broken open in a single move. En passant still applies, wherever the double step happened.
Sidestep Pawns
Pawns may also step one square sideways, without capturing. It sounds small and changes everything: doubled pawns can untangle, a blocked chain can shuffle out of the way, and files open and close constantly. Nothing stays fixed, so the pawn structure stops being the stable part of the position.
Extinction Chess
Lose every piece of one kind and you lose the game — R. Wayne Schmittberger's 1985 idea. You start with a single queen and a single king, so losing either ends it at once and a queen trade is a resignation. There is no check: the king is just another piece. Promoting your last pawn loses too, because your pawns then hit zero.
Upside-down Chess
Both armies start at the wrong end of the board, so each side's pawns sit in front of its own back rank and face the wrong way. Only the knights can move at all on move one — everything else is boxed in by its own pawns. Every opening book you have ever read is worthless from the first move.
How strong is the AI?
Honest answer: the variant AI is starting strength, and we expect to improve it.
Standard chess in Chess Ta! runs a trained NNUE neural network, fitted on nearly two million unique positions from the engine's own self-play. The seventeen variants do not have that yet. They currently play on the engine's hand-written evaluation — real search, real tactics, but no learned network behind it.
That is a deliberate choice rather than an oversight. In this engine a network replaces the evaluation instead of correcting it, so a network that has not learned enough plays worse than no network at all. We trained one for each of the newest variants and measured them against the hand-written evaluation in real games — and every one of them lost. So none of them shipped. Training on millions of games per variant is what would change that, and it is planned rather than done.
Until then the variants play a solid, honest game and will not embarrass themselves, but they are not yet as strong as the standard chess engine.
Built from scratch — no Stockfish, no borrowed engine
The engine underneath all of this is ours. It is written in C with no third-party dependencies: the bitboard representation, magic-bitboard sliding attacks, move generation, the alpha-beta search with its transposition table and pruning, the evaluation, the NNUE inference, and the Go trainer that produces the networks. There is no Stockfish in it, no chess library doing the thinking, and nothing calling out to a server.
Each variant is a small rule module plugged into that shared engine, which is why seventeen of them fit in an app that still runs offline on a phone. If you want the details of how the search and the training pipeline actually work, they are written up on the engine page.
Questions
Does it work offline?
Yes. The engine is compiled into the app and runs on your device. Nothing is sent to a server, no account is needed, and every variant is playable with no connection at all.
Is this based on Stockfish?
No. The engine is written from scratch in C — the board representation, move generation, search, evaluation and the neural-network code are all ours, with no third-party engine and no external dependencies.
How strong is the variant AI?
It is starting strength. Standard chess runs a trained neural network, but the seventeen variants currently use the hand-written evaluation. Their networks have not been trained on millions of games yet, and we expect to improve them in a future update.
Which variants support Chess960?
Most of them. Any variant whose armies start from a standard back rank can be shuffled into a Chess960 starting position, which covers most of the list including all five added in version 1.4.0.
Do I have to pay for the variants separately?
No. Chess Ta! is a paid app on the Microsoft Store — a single one-time purchase, with no subscription. All seventeen variants are included in it, so there is nothing further to unlock before you can play any of them.