miHoYo releases open-source AI models for game environment interaction
A research team backed by miHoYo has released GameScaling, a suite of vision-language models capable of controlling video games through keyboard and mouse inputs.
A research group including scientists from miHoYo and university partners has made the GameScaling model family available to the public. These models are designed to interpret visual data from game environments and output corresponding hardware inputs, allowing AI agents to navigate and interact with various digital landscapes.
The project encompasses eight distinct model checkpoints, ranging in size from 800 million to 27 billion parameters. These iterations are built upon the architecture of Qwen3.5 vision-language models. By training on gameplay footage, the models have demonstrated the ability to perform complex actions such as resource gathering in Minecraft, tactical combat in Doom, and navigation within puzzle-based environments.
Released under an Apache 2.0 license, the initiative includes model weights, inference code, and sample programs. The release aims to provide researchers with tools to examine how training across diverse virtual environments influences the performance and adaptability of interactive AI agents.