ashtonanderson's Blog • Introducing Maia-3: free and open source • lichess.org<br>Accessibility - Enable blind modeYour network blocks the Lichess assets!<br>Donate

Sign inRegister

Introducing Maia-3: free and open source<br>ashtonanderson24 May 202624411,703 viewsEnglish (US)<br>Chess engineChessSoftware Development<br>A human-like chess engine open to all Today we are releasing and open-sourcing Maia-3 , our latest model of human behavior in chess! Maia's goal is to predict the human move, not necessarily the best move, in any chess position.

The Maia bots have been playing millions of games on Lichess for years. They're here to stay, but the Maia models that power them can do so much more. Maia-3 now models players from 600 to 2600, so it can power applications that apply to beginners and titled players alike. Maia-3 is the new state of the art, predicting the move 57.1% of the time on a standard test set of positions, compared to 52.0% for Maia-2 and 51.6% for Maia-1. ALLIE, another recent model, achieves 55.7% but needed an architecture almost 10x as large, making Maia-3 more efficient while still being more accurate.

For builders, download Maia-3 here, use the source code here, and read the research paper here.

If you want to use Maia-3 to play, analyze, drill openings and endgames, do tactics, and more, go to www.maiachess.com.

Join our Discord to connect with our community. We'd love to hear what you think and see what you build using Maia-3!

Maia-3 helps you enjoy and understand chess through a human lens. It lets you see not only the engine move, but which moves real players see, which mistakes they’re drawn toward, and how that changes the way you play, analyze, improve, and have fun with the game. Maia-3 unlocks lots of possibilities: human evaluation bars, human-like bots, new kinds of tactics puzzles, coaching aids, position difficulty measurement, and more. We can't wait to see what the community builds with it!

Features

Maia-3 has lots of benefits:

You can play against it for free on www.maiachess.com. It's a training partner that plays at your choice of skill level, is the most human-like engine out there, never gets tired, and is a low-stakes way to get some chess in.

Maia-3 now models players rated from 600 to 2600 (Lichess blitz ratings), so it spans 99% of the skill spectrum. Tools you build with it can apply to beginners or titled players.

Maia-3 powers the www.maiachess.com platform, where you can do a new kind of analysis (combining the strength of Stockfish and the human side of Maia-3), play Hand and Brain, drill openings and endgames (pick some openings and endgames you want to practice, a skill level to train against, and go), try your hand at our chess Turing test called Bot-or-not (can you tell if a human or machine is playing?), follow tournament broadcasts and any Lichess game live, and more. For an example of Maia-3 powered analysis, see below.

It is designed to be interpretable, meaning it sees things on the chessboard similarly to how we see them. This makes it a lot easier to build educational and fun tools on top of Maia-3.

Chessformer, our new transformer architecture for chess modeling

Maia-3 uses a completely new architecture called Chessformer. Concretely, Chessformer is an encoder-only transformer that treats the 64 board squares as tokens, pairs this square-token body with an attention-based “source-destination” policy head, and equips the trunk with Geometric Attention Bias (GAB), a novel dynamic positional-bias generator that adapts attention to the geometry of a chess position.

Chessformer is cool because it let us achieve the best human emulation performance, but we also integrated it into Leela Chess Zero and it resulted in a 100 Elo point gain for optimal play when search is turned off, plus it is interpretable by design, meaning it's a lot easier to figure out Maia-3's "thoughts". Prior to this, if you wanted to emulate human play you used one model (like Maia-2), but if you wanted raw engine strength you used another architecture (like Stockfish), but if you wanted something understandable you had to use a third strategy (like hand-crafted evaluations). Now you can just use Chessformer for everything. For those interested in all of the details, you can read our paper "Chessformer: A Unified Architecture for Chess Modeling", which was just published in ICLR 2026 (a top machine learning conference).

Analyze games with Maia

As an example of what Maia-3 enables, try Maia-3 analysis on www.maiachess.com. You can analyze any game—your Lichess games, famous historical games, or any game or position you want—with a dual Maia-3 / Stockfish analysis view. By augmenting traditional engine analysis with Maia-3's human understanding, you get real-world context in every position. In Maia analysis, you see:

What Maia predicts people at different rating levels would actually play, and with what probability

What Stockfish thinks is objectively best

A "Moves by Rating"...