chess-gpt-uci-square (C)

A 50.6M-parameter GPT that plays chess by predicting the next move token. There is no search and no evaluation function: one forward pass per move. It is model C of three models trained the same way and differing only in their tokenizer:

Model	Tokenizer	A token is	Vocab
A	pgn-char	one character of 1.e4 e5 2.Nf3	34
B	uci-move	one move, g1f3	1971
C	uci-square	one square (+ promotion piece), g1 f3	71

Write-up: blog.wezn.dev/chess-gpt · Code: github.com/wezn/chess-gpt

This model

• Architecture: decoder-only transformer, 16 layers, 8 heads, width 512, context 384 tokens, tied embeddings
• Training data: 35M Lichess games (rapid and slower, all ratings), each seen once
• Training: 76,110 steps on one NVIDIA GB10
• Validation loss per move: 1.604

Mean over the final 11 snapshots, from the write-up:

	A	B	C
Elo (Lichess scale, estimated from self-play scored by Stockfish)	1,940	1,908	1,792
Legal moves on held-out human games	99.74%	99.77%	99.84%
Average centipawn loss	56	58	61

The Elo figures are estimates, not results from rated play. The write-up describes how they were measured.

Usage

git clone https://github.com/wezn/chess-gpt && cd chess-gpt && uv sync
uv run hf download ftaz/chess-gpt-uci-square --local-dir models/chess-gpt-uci-square

import chess
from chessgpt.eval.player import ModelPlayer

player = ModelPlayer("models/chess-gpt-uci-square", device="cpu")
board = chess.Board()
for san in ["e4", "e5", "Nf3", "Nc6", "Bb5", "a6"]:
    board.push_san(san)
print(player.choose_moves([board])[0].move)                         # the model's move
print([(board.san(m), round(p, 3)) for m, p in player.rank_legal_moves(board)[:3]])

config.json holds the architecture and tokenizer. model.safetensors holds the weights only.