The Maia Chess team has officially launched a new open-source chess engine called "Maia 3." Trained on 250 million real human games, the model has achieved an Elo rating of approximately 1800, representing a nearly 300-point leap from the previous generation. More importantly, this engine is completely free and open-source, supporting local deployment, marking a key step toward making AI chess engines more accessible to the general public.

Unique Approach: Focusing on Simulating Human Decision-Making

Differing from traditional top-tier engines like AlphaZero and Stockfish, which aim for superhuman levels with Elo ratings over 3500, the core purpose of the Maia project is to explore how AI can better understand and simulate human chess behavior. It focuses on "playing like a human"—predicting actual human moves rather than seeking absolute mathematical optimal solutions.

This "human-centered" training approach ensures that Maia 3 does not make moves that are difficult for humans to understand—instead, it accurately simulates common patterns, psychological preferences, and even logical errors found in real games. This makes it capable of overcoming the frustration caused by strong engines, becoming a more educational and valuable companion for amateur club-level players.

Llama Architecture and Expansion to Multiple Chess Variants

In terms of technical architecture, Maia 3 has undergone significant improvements. It is built based on Meta's Llama 3.1 architecture, using a decoder-only transformer design. Compared to the previous GPT-2-based architecture, this offers stronger context understanding and computational efficiency. Its training data covers various skill levels, from beginners to experts.

Notably, Maia 3 is no longer limited to chess alone but has expanded to support multiple board games, including Shogi (Japanese chess), Go, and Xiangqi (Chinese chess). This evolution into a multimodal capability has made it a versatile platform for general chess-related AI academic research.

Smooth Operation on Consumer Hardware Promotes AI Democratization

Maia 3 is released under the highly friendly Apache 2.0 license, meaning developers worldwide can freely use, modify, distribute, and even commercialize it. Currently, the project's code, training data, and model weights have been fully released on GitHub without any reservations.

To significantly lower the entry barrier, the team has performed deep hardware optimization on Maia 3. The official statement says that users do not need high-end graphics cards; a regular laptop with a modern CPU can run it smoothly in local deployment. Players can now directly play against Maia 3 on the lichess.org platform, and the engine will provide high-quality free underlying support for future online game platforms and teaching software.