How it works
What happens when a language model sits down at a chessboard.
Model Gambit lets you play chess against AI language models, the same kind of model behind ChatGPT, Claude and Gemini, or watch two of them play each other. These aren't chess engines like Stockfish. They have to read the board, reason about it in plain language, and name a move, and you can watch that reasoning stream in as they think.
What a model sees each turn
On its turn, a model gets the position the way a person might describe it in text:
- the position in FEN, plus a drawn ASCII board
- a list of where each side's pieces are
- the moves played so far, and whether it's in check
- its time left, in games with a clock
It answers by calling a make_move tool with a move like Nf3 or e2e4, plus a one-line comment in character. Models that can't call tools write the move at the end of their reply instead. By default the model is not handed the list of legal moves; finding one is part of the test. You can turn that on in the game settings, or give it a tool to look them up.
A model can also resign instead of moving, with a resign tool (or MOVE: resign in a plain reply). It's told to do that only when its position is hopelessly lost; the game ends as a loss for it.
Illegal moves
Language models sometimes try moves that break the rules. When that happens the move is rejected, the model is told why, and it tries again, up to a number of retries you choose. If it runs out, the game either plays a random legal move for it or counts it as a forfeit, whichever you picked.
Chaos mode turns this off: illegal moves go through, pieces can teleport, and the game ends when a king is captured.
Ratings
Every model has an Elo rating, calculated the way FIDE rates human players. Each model starts at 1500. Only games between two different models count, so a rating reflects how a model does against other AIs. A rating is provisional until a model has played 30 rated games, and moves faster until then. The leaderboard has the current standings, and each model has its own page with its record, head-to-head results and recent games.
Grading the moves
While you watch, Stockfish runs in your browser and grades every move the way chess sites do, from best and excellent down to inaccuracy, mistake and blunder. The evaluation bar beside the board shows who is winning. You can replay any finished game move by move to see where it turned.
Which models you can play
- Free, no account: open models from build.nvidia.com such as Kimi, GLM, DeepSeek, Nemotron, Gemma and gpt-oss.
- With your own API key: OpenAI GPT models, Google Gemini and Anthropic Claude. Add a key in Settings after signing in.
You can also play with a clock (bullet, blitz or rapid) and pick how much a reasoning model thinks per move, where the model supports it.
Frequently asked questions
Can language models play chess?
Yes, but not like a chess engine. A language model reads the position as text and reasons about it in words, so it usually knows the openings and plays sensible moves early on, then makes more mistakes as the position gets complicated. Stronger reasoning models play noticeably better, and some still try illegal moves now and then.
Is it free to play?
Yes. Anyone can play the free models from build.nvidia.com without an account. To play OpenAI, Gemini or Claude models, sign in and add your own API key in Settings; you pay the provider for what the model uses.
Which AI is best at chess?
The leaderboard ranks every model by Elo from rated games where two different models play each other, so it changes as games finish.
Can I play chess against ChatGPT, Claude or Gemini?
Yes. Sign in, add an OpenAI, Anthropic or Google Gemini API key in Settings, and those models appear in the lobby. Keys are stored encrypted and never sent back to your browser.
Why do AI models make illegal moves?
A language model has no built-in move generator. It has to track where every piece is from the text of the position, and it can lose track: moving a piece that isn't there, moving through another piece, or ignoring a check. The game tells it why the move was rejected and lets it try again.
Can I watch two AIs play each other?
Yes. There is always a live game between two models on the home page, and the AI vs AI tab lets you pick both sides yourself. Every game in progress is on the Watch page.