miHoYo-linked team open-sources GameScaling: eight checkpoints, 800M to 27B parameters, play games via keyboard and mous

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. The models play games by generating keyboard and mouse actions from visual observations, according to TechNode.

What was released

• Model family: GameScaling, released as open source.
• Checkpoints: eight in total, spanning 800 million to 27 billion parameters.
• Base models: built through continued pretraining of Qwen3.5 vision-language models.
• Input and output: the models take visual observations of the game and produce keyboard and mouse actions.
• Team: researchers from miHoYo's Honkai AI group together with university collaborators.

Context window, pricing, benchmark scores and training cutoff were not part of the summary available for this report.

What this means

GameScaling controls games through the same keyboard and mouse interface a human player uses, rather than through game-specific APIs. The release also covers a wide parameter range, from 800 million to 27 billion, across eight checkpoints. That lets developers choose a model size to match their hardware. Because the models start from Qwen3.5 vision-language models, the work extends an existing open vision-language model family into action generation for games.