CSSLab

CSSLab / maia3

Public

Maia-3 is the most accurate and efficient human chess move prediction engine.

47
5
100% credibility
Found May 26, 2026 at 47 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

Maia-3 is a chess artificial intelligence created by researchers at the University of Toronto that predicts how human players of different skill levels would move in any position. Unlike traditional chess engines that play 'perfectly,' Maia-3 is designed to make human-like mistakes—it can imitate a 1000 Elo player, a 1500 Elo player, or any other skill level. The project provides several AI models of different sizes that you can run as a chess engine in any chess program. You can adjust settings like the temperature to control how creative or conservative the AI plays, and it will analyze positions showing you what moves a human at your chosen level would likely make. It's particularly useful for players who want to practice against opponents that match their skill level, analyze games to find where human-like errors occurred, or train by playing against realistic human behavior rather than computer-perfect play.

How It Works

1
♟️ You discover Maia-3

You hear about a chess AI that doesn't just play perfectly—it plays like a human at your skill level.

2
📦 You install the package

With one simple command, the chess engine installs on your computer and everything is ready to go.

3
⬇️ The model downloads automatically

The first time you run it, the AI brain downloads from the internet and saves to your computer for next time.

4
🎮 You open your chess program

You add Maia-3 as an opponent in any chess program you already use—it works with popular chess apps.

5
You choose how to play
🐢
Quick games on your computer

Use the smaller model that runs fast on any machine

🚀
Best accuracy for analysis

Use the larger model for more human-like predictions

6
🤖 You play against human-like moves

The AI makes moves that feel natural for the skill level you chose—no perfect play, just realistic human mistakes.

🏆 You have a human-like training partner

Whether practicing tactics, analyzing your games, or just having fun, you now have an AI that thinks like a human opponent.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 47 to 47 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is maia3?

Maia-3 is a chess engine that predicts what a human would play, rather than finding the objectively best move. Built in Python with a transformer architecture called Chessformer, it plugs into any UCI-compatible chess GUI just like Stockfish or Lc0. You set an Elo rating, and it outputs moves that match how players at that skill level actually think. Four model sizes are available, ranging from 3 million to 79 million parameters, with the larger models offering better accuracy at the cost of more compute.

Why is it gaining traction?

The hook is the use case: analyzing positions the way a human 1500-rated player would, not the way an engine would. This makes it useful for training tools, post-game analysis that feels instructive rather than intimidating, and studying common mistakes at specific rating levels. The CLI is straightforward -- `maia3-5m` to run, `--elo 1200` to target a skill level, `--multipv 5` to see the top human-like alternatives. Checkpoints download automatically from Hugging Face, and the engine speaks standard UCI, so it works with Lichess, Nibbler, or any desktop GUI out of the box.

Who should use this?

Chess coaches building training positions that match their students' skill levels. Developers building analysis tools who want human-readable evaluations instead of centipawn soup. Researchers studying how players at different ratings diverge from optimal play. If you want an engine that plays like a human opponent rather than crushing you with superhuman precision, this is the tool.

Verdict

With only 47 stars and a 1.0% credibility score, this is a young project backed by an academic research lab. The paper is recent, the implementation is clean, and the UCI integration works -- but test coverage and production hardening are unknown. Worth exploring for the human-move prediction angle, but treat it as research-grade software rather than a drop-in replacement for established engines.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.