HN Showcase
Back to browse
GitHub Repository
0 starsPython

Chess bot based on the transformer architecture

by tchauffi·Jun 17, 2026·3 points·0 comments
Visit ProjectView on HN

AI Analysis

●●SolidBig BrainNiche Gem

Human-only training with custom chess-geometry attention bias reaches 2100 Elo.

Strengths
  • •Custom chess-geometry relative bias in attention is a genuinely novel architectural choice.
  • •MCTS harness design extracting max performance from small model shows real engineering insight.
  • •Live Lichess bot allows real-world evaluation against human players.
Weaknesses
  • •Chess engine space is dominated by Stockfish and Leela with vastly superior performance.
  • •11M parameters is tiny but 2100 Elo won't attract serious players.
Category
AI/ML
Target Audience

ML researchers and chess enthusiasts

Post Description

Hi HN!

I build this project to explore an idea I got in mind for a long time : Is transformer a suitable architecture for a chess bot? I built a small model (11M parameters) and trained it on human games (Elite Lichess DB).

Model alone is performing around 1500 elo, but I built an harness using Monte Carlos Tree Search (MCTS) using my model heuristics to improve the model to ~2100 elo (evaluated against stockfish).

If you want to try it, it is available as a Lichess bot : https://lichess.org/@/ChessTransformerBot

I'm looking to evaluate this model against human players so challenge, I would be grateful if you try it!

The project is open source, don't hesitate to star the repos if you like the project.

For me, the main key learning is that machine learning is an important part of the project, but it was the harness design that makes the system works with a nice performance regarding the small model size.

Similar Projects