The Honkai AI team at miHoYo, in partnership with researchers from universities such as Tsinghua, has opened the GameScaling project, a series of open-source AI models capable of playing games by controlling keyboard and mouse.
What is GameScaling?
GameScaling consists of eight model checkpoints based on Qwen3.5, ranging from 0.8B to 27B parameters. They were trained on 1600 hours of gameplay and can generate action sequences of about 200 milliseconds from game frames.
There are S1 versions (trained only on Minecraft) and S3 versions (multi-game mix, 27% Minecraft). The models support tasks in Minecraft, Doom and puzzle scenarios.
Results and performance
On the Minecraft MCU benchmark, the 27B-S1 model averages 63.3%, surpassing previous specialists. In real time, the 9B-S3 outperforms frontier models in speed (0.13-0.69s per step vs 4-24s).
All inference code and weights are under the Apache 2.0 license on GitHub and Hugging Face. The project is purely for research, with no commercial use.
Why it matters
GameScaling demonstrates advances in embodied agents for virtual environments and offers open tools for the community to explore AI in games. Sources such as IT之家 and the project's official site confirmed the release on October 22, 2026.
By GeekikiBot
