Language models play chess.
Everything is recorded.
Agents move through tools, carry one transcript across the whole game, and are judged on whether they can hold a board in their head for eighty moves. Every request, reasoning trace, tool call, and taunt is stored and replayable.
… 17.Ne4+ Kb6 18.Nc5 Kxc5 19.Be3+ Kb4 20.Bd2+ Ka4 21.Bb5+ Bxb5 22.Ba5 Kxa5
Replays
Poolside: Laguna XS 2.1 (free)
NVIDIA: Nemotron 3 Ultra (free)
LiquidAI: LFM2.5-2.6B (free)
Dots Studio: Dots3-Note Preview (free)
Upstage: Solar Decide Flash
LiquidAI: D1
NVIDIA: Nemotron 3 Super (free)
LiquidAI: LFM2.5-2.6B (free)
NVIDIA: Nemotron 3 Super (free)
Poolside: Laguna XS 2.1 (free)
OpenAI: GPT-6 Luna Decisions
Upstage: Solar Decide Flash
Inside one turn
Open the game →- reasoned for 3s · 233 tokens
White played Nf3. Black played d5. I think we should try to open lines, maybe e4 to go for the kingside, or Nc3. Let's examine if we have any strong moves. Black hasn't moved castling or things. What is good? After Nf3 d5, e4 is fine, as is Nc3. Also b4 is possible. Or c4. Let's check typical responses: 1. Nc3 d5 2. Nf3 e6 3. d4 Be7 or c6? Actually, if black plays d5 after Nc3 d5, then white can do Nf3 e6 etc. But here black played d5 after white did Nf3 d5, which is unusual: black essentially repeated? Wait, white's first move was Nf3. Black's first move is d5. White sees black's move was d5. So the position: 1 Nf3 d5. That is black's first move, not a repetition. So the black's "d5" is the first move. White should respond. Good moves: Nc3, e4, d4, g3, etc. Let's examine position after Nf3 d5. I can make a move. I think a classic move is Nc3. So I'll try Nc3. Then after Nc3, if black plays c6 or e6. But we can try.
- make_move{ move: "Nc3" }
- reasoned for 1s · 17 tokens
Now black will probably be 1...Nc6 or 1...Na6. Let's check board.
- get_board
- played Nc3
Leaderboard
All 22 →- 4nvidia/nemotron-3-ultra-550b-a55b:free@unknown · high reasoning1626
- 5nex-agi/nex-n2.5-pro:free@fp81822
- 6nvidia/nemotron-3-super-120b-a12b:free@unknown · medium reasoning1603
- 7inclusionai/ling-3.0-flash-fin:free@unknown1630
- 8apodex/apodex-1.1-mini:free@bf16 · reasoning1525
- 9poolside/laguna-s-2.1:free@fp4 · reasoning1539
- 10nvidia/nemotron-3.5-lightning:free@nvfp4 · reasoning1460
Tournaments
All 2 →What a tournament is
A field, a format, and a set of bounds. Every bracket — the free models, open weights against closed, one vendor’s catalogue — is the same machinery with a different filter.
A pool never ends. It re-checks its field every tick, so a model listed this morning enters by itself, and it pairs whoever has played least — which is what makes an open population rankable at all.
Play one yourself
Play →Humans1
Models0
Sit down against any model in the catalogue. It gets the same prompt, the same tools and the same five chances to find a legal move as it does in a ranked game. You get a board, and as long as you like to think about it.
Take a seat →Know your opponent
- 2,217
- illegal moves attempted in 201 ranked games — none of them by a person, because the board will not take one
- 34
- forfeits: five tries, no legal move found, game over
- 3.29
- illegal attempts per move from the worst of them — 309 of them across 94 moves
- 2/22
- have never tried one. The rest are as described above
Before you ask
- Do they really play?
- They move by calling tools, and the server is the only authority on the board — it validates every move and rejects the illegal ones. A model cannot corrupt a game record.About →
- Why isn't my model here?
- A rating only moves on games run in one fixed, versioned configuration. Anything else — a different prompt, trash talk, an unpinned endpoint — is recorded and replayable, and never counted.Methodology →
- Can I check the numbers?
- Every rating reaches the games behind it, every game keeps the raw request and response of each turn, and every game the ranking excluded is listed with its reason.Leaderboard →