We are pleased to announce the release of the Chess LLM Baselines for the Global Chess Challenge 2025 
. This baseline provides a complete supervised fine-tuning pipeline for training language models on chess, centered around a special-token encoding of board positions and legal moves 
. By assigning each square and piece its own unambiguous token, the approach avoids common tokenization issues associated with FEN strings while remaining significantly more compact than ASCII board representations, leading to faster convergence and a substantial reduction in illegal move generation 
.
The baseline includes fully supported training workflows for AWS Trainium (Neuron) and NVIDIA GPUs, enabling researchers and practitioners to run experiments across diverse hardware environments. While the reference implementation fine-tunes Qwen3-0.6B, the methodology is deliberately model- and architecture-agnostic 
. These are the weights for a model trained on Qwen3-4B. Any causal language model and model size can be used with the same encoding scheme, tokenizer setup, and training logic, allowing straightforward scaling or architectural experimentation.
In addition to training, the release provides an integrated evaluation framework covering puzzle solving 
, games against a random opponent 
, and games against a weak Stockfish configuration 
. Together, these components offer a robust and extensible foundation for developing chess-playing language models, from rapid prototyping to more ambitious competitive submissions.
All the best for the remaining days of the challenge.