Mobile Maia

Maia is a chess engine from the University of Toronto’s Computational Social Science Lab that tries to play like a human, not like the strongest engine. In the project’s words: “Maia is a human-like chess engine, designed to play like a human instead of playing the strongest moves.” It is trained on games from Lichess to predict the move a player of a given rating would make, mistakes included. Mobile Maia puts the newest version, Maia-3, into an Android app. Its README: “Mobile Maia is a community project for playing and studying chess with Maia on Android. It will always be free, never show ads, and always remain open source.”

What you can do

• Play Maia at a chosen rating from 600 to 2600. The README: “Maia-3 runs on-device and selects moves for the chosen rating instead of behaving like a deliberately weakened tactical engine.” Games can be untimed or use Lichess-style time controls, from 1+0 to 30+20, or a custom one.
• Analyse positions on a board that shows two lines from a light build of Stockfish 19 next to the moves Maia expects players of one or two ratings to choose. Version 2.5 can also show Maia’s expected-score percentage, including draws, for one or two ratings; it is optional and the app links to the developer’s reliability study rather than presenting it as a certain result.
• Review a whole game, with each move marked from Brilliant to Blunder, “without a daily quota”.
• Import and export games as PGN and positions as FEN, with opening names from Lichess’s opening list, also offline. A graphical editor in 2.5 can place, move and remove pieces and copy or share the resulting FEN.
• Play on a Chessnut e-board over Bluetooth, an experimental feature confirmed with the Chessnut Go and Chessnut Air; timed games on the board aren’t supported yet.
• Change the appearance with 25 board designs and 16 piece sets. The app theme can be light, dark or follow the system, with separate options for system colours and a pure-black background. Existing installations keep their saved choices.

The model it uses, Maia-3 with 79 million parameters, was trained on Lichess games from January 2023 to July 2025. Its model card is clear about its purpose: “Not intended for maximum playing strength.”

About this file

• This is version 2.5.1. Version 2.5.0 added the appearance controls, graphical position editor, optional Maia score, continuation from an analysed position on a Chessnut board and smoother variation navigation; 2.5.1 fixes a crash when analysing certain custom positions. Version 2.4.0 added opening-variety and sampling-data options. In 2.5, new or reset settings start with opening variety on and sampling capture off, with Maia 1500, the second engine at 2600, Temperature 1.0 and Top-P 0.95 as the other new defaults; existing saved choices remain unchanged. Version 2.3.0 added ten interface languages and Lichess-style quick time controls.
• The file is F-Droid’s, which builds the app from source, checks that its build matches the developer’s, and then distributes the developer’s own signed APK: it is byte for byte the release on GitHub, signed with the key whose fingerprint the README publishes (cd6c07c4…f172e). There is no Google Play version. A separate Mobile Maia Preview app, with a gold icon, installs beside it.
• The 455.4 MB file runs on 64-bit ARM, 32-bit ARM and 64-bit x86, and needs Android 7.0 or later. It declares 6 permissions: five for the Chessnut board, Bluetooth and, on Android 11 and older, location, which Android then required for Bluetooth scanning, plus one AndroidX adds for internal use. Our scan found no trackers.
• The app’s code is licensed under the AGPL-3.0, and so are the Maia-3 model weights; Stockfish and the Lichess sounds and chess libraries it uses are under the GPL-3.0. Version 2.5’s source records a rights review for the piece art: the APK contains the 16 sets offered by the selector, while 24 non-commercial or unclear-rights sets from the underlying board library are excluded from the package.
• The developer is a GitHub user called Dash; the app is a community project and isn’t affiliated with Lichess or the University of Toronto.
• The screenshots are the developer’s, from F-Droid, taken with version 2.2.

CPU types in this build: arm64-v8a (64-bit ARM, almost every phone since 2017); armeabi-v7a (32-bit ARM, older and budget phones); x86_64 (64-bit Intel, emulators and Chromebooks).

Versions currently kept in the F-Droid repository. We host and check only the newest one. Full changelog.