MiHoYo team open-sources GameScaling models that control games through keyboard and mouse actions

A research team involving miHoYo's Honkai AI researchers and university collaborators has open-sourced GameScaling, a family of eight checkpoints ranging from 800 million to 27 billion parameters that play games by generating keyboard and mouse actions from visual observations. The models were built through continued pretraining of Qwen3.5 vision-language models and use current and recent game frames plus previous actions to produce short control sequences, demonstrated on mining and crafting in Minecraft, combat and navigation in Doom, and puzzle-solving in escape-room environments. The project releases model weights, inference code and example programs under the Apache 2.0 license, letting developers study how training across different virtual worlds affects interactive AI performance.