The Short Answer
Quick answer: Maia is a chess engine trained to play like a human, not to win. Researchers at the University of Toronto taught it on millions of Lichess games, so it picks the move a player at a given rating would likely make. You can play its three featured bots on Lichess for free.
I read the Maia project's README on GitHub and the research papers it links. Then I checked the Maia website and the Lichess bot list on October 18, 2026.
Here's how Maia works, how strong it really is, and whether you should play it.
What Maia Is
Maia is a chess engine with an unusual goal. It doesn't try to find the best move. It tries to find the move a human would play.
The project comes from a research team at the University of Toronto, led by Ashton Anderson. The Maia website lists Jon Kleinberg of Cornell University and Siddhartha Sen of Microsoft Research as collaborators.
Its homepage puts the idea simply. Maia is trained to play like a human rather than to win. It adds that Maia shows common human biases and makes many of the same mistakes that humans make.
That's the whole point. For practice, an opponent that slips the way your rivals slip can matter more than raw strength.
Why Normal Engines Don't Play Like People
Stockfish and other strong engines are built to win. Maia's researchers asked a different question: how well do engines predict what you or I would actually play?
The first Maia paper, which appeared at KDD 2020, tested that. Its authors ran existing engines, including an open source version of AlphaZero, on online games and checked how often each one matched the human's real move. Their finding was blunt. Those engines don't predict human moves well.
Maia did much better. The paper says it predicts human moves at a much higher accuracy than existing engines.
That's why a weakened engine can feel odd when you play it. Its mistakes come from limits placed on a strong engine, not from the habits of a real player at your level.
How Maia Learned From Human Games
Maia learned from people, not from playing against itself.
The paper calls it a customized version of AlphaZero, trained on human chess games. On GitHub, the Maia code stores the final models as Leela Chess neural networks.
The games came from Lichess. Its bot list says each of the three featured Maia bots learned from over 10 million Lichess games between players around one rating, either 1100, 1500 or 1900.
So maia1 studied a huge number of positions from players near 1100, along with the moves they chose. It learned which moves those players tend to pick, including the mistakes you'd expect from them.
The team thanks Lichess on its website for providing the human games it trained on.
Nine Models, From 1100 to 1900
The first version of Maia isn't one bot. It's a set of nine.
The GitHub page describes them as a collection of chess engines that play like humans, from 1100 to 1900. There's one model for each 100 point step: 1100, 1200, 1300 and so on, up to 1900.
Each model learned only from games around its own rating. The paper says Maia reaches its best accuracy when it predicts moves by players at a specific skill level, and that you can tune it to that level.
In practice, you pick the model nearest your own rating, and it plays the kind of moves your usual opponents play. Just remember that these are Lichess ratings, so your rating on another site may not match.
It Barely Calculates
Here's the strangest part. Maia hardly searches at all.
Most engines look through a huge number of positions before they move. The Maia README says to switch that search off. The team runs the models with a limit of one node, which means the engine looks at the current position and picks a move straight away.
So Maia plays more like a quick glance than a long think. It picks the move its network judges a player at that rating would most likely make.
That makes it fast. It also means Maia can miss a tactic that you'd find with a few moves of calculation. For a bot meant to copy club players, that's a feature.
Why Maia Plays Above Its Target
If you're rated 1100, don't expect maia1 to play like you on a bad day.
The README says the models are stronger than the rating they're trained on, because they make the average move of a player at that rating. A real 1100 player has good and bad moments. Maia smooths those out.
The Lichess numbers show it. When I opened the Lichess bot list on October 18, 2026, maia1, which targets 1100, had a rapid rating of 1655. Its classical rating was 1647, and its blitz rating was 1422.
So treat the target as a style, not a strength. Maia1 makes the choices an 1100 player tends to make, but more steadily, and its rating sits well above 1100.
The Three Maia Bots on Lichess
Lichess lists three Maia bots among its featured bots, with a Play button next to each one.
- Maia1 targets 1100. Its rapid rating was 1655 over 641,262 games.
- Maia5 targets 1500. It stood at 1704 in rapid, over 234,402 games.
- Maia9 targets 1900. Its rapid rating was 1750 over 219,358 games.
There's also MaiaMystery, which the README says is for testing new versions of Maia. A Lichess team called maia-bots exists too, and the README says more bots will be added to it.
Look at the gap. Maia1 and maia9 target ratings 800 points apart, but their rapid ratings were less than 100 points apart that day.
Same Position, Same Move
Maia has one quirk you'll notice if you play it often.
The README says the models play the same move every time. That's why the Lichess bots use opening books, which the README describes as still in development.
Without a book, many of your games against maia5 could start the same way. The book adds some variety to the first few moves.
You can use this quirk too. If you find a line that beats a Maia bot, it may walk into it again. That's handy for drilling one opening, but it doesn't prove the line works against a human, who will vary their play from game to game in ways a fixed model won't.
Maia-2: One Model for Every Level
In 2024, the team released Maia-2, presented at NeurIPS that year.
Its big change is one model instead of nine. The Maia-2 paper says earlier work used completely independent models for each skill level, so those models couldn't follow you smoothly as you improved.
Maia-2 takes your skill into account directly. The paper describes a skill aware attention mechanism that mixes your strength with the position on the board.
Its training data came from rated rapid and blitz games on Lichess. The released training range runs from monthly archives of May 2018 to November 2023, and the Maia-2 README says games involving bot accounts are left out.
Maia-3 and Chessformer
Maia-3 is the newest version, and the GitHub page announces it as now available.
It's built on a new design called Chessformer, described in a paper presented at ICML 2026. The paper says Maia-3 reaches 57.1 percent move matching accuracy. In other words, it guessed the human's real move in about 57 of every 100 test positions.
It also beat the previous best result with fewer than a quarter of the parameters, a common measure of model size.
Chessformer helps the strong engines you already know, too. The same paper says adding it to Leela Chess Zero gave over 100 Elo of extra playing strength. Our comparison of Stockfish, Leela and Komodo explains where Leela fits.
What the Maia Website Offers
The Maia website is more than a project page. You sign in with your Lichess account, and it has its own tools.
The menu lists play, analysis, puzzles, Hand and Brain, opening drills, and a game called Bot or Not, where you guess whether moves came from a human or an AI. The site says Maia can target any rating on the Lichess scale, from 600 to 2600.
It pitches Maia as a tool for training, game analysis and coaching. The idea is that you learn from realistic decisions rather than computer perfection.
The team also studies personal models. One of its papers describes bots that mirror a single player's style from just 20 of their games.
Who Should Play Maia
Maia suits a player who wants practice games that feel like real games.
If you want an opponent that makes the kind of choices your rivals make, Maia is built for exactly that. Its mistakes are meant to follow human patterns, so the lessons should carry over to games against people.
It's less useful if you want a ladder of fixed strengths. The three Lichess bots sit close together in rating, and none of them plays near master strength.
For a wider choice of computer opponents, our guide to Chess.com bots vs the Lichess computer compares both sites. If bots around 1500 keep beating you, read how to beat 1500 rated chess bots.
How to Learn From Maia Games
A game against Maia only helps if you look at it afterwards.
After each game, find the moment where the result turned. Ask whether you'd have spotted the same chance against a person. Maia's slips are meant to look like human ones, so the patterns you find there may well repeat in your own rated games.
Then fix one thing at a time. Keep losing in the opening? Drill that line before your next session.
The CheckmateX bot trainer lets you set a strength from 800 to 2800, pick a side, and start from a chosen opening line. That's a handy way to rehearse the same position again and again.