Skip to main content
Back

miHoYo Drops Eight Open GameScaling Models That Watch and Play PC Games

miHoYo and Tsinghua release GameScaling, a family of vision-language models that play keyboard-and-mouse games in 200 ms action chunks.

  • miHoYo and Tsinghua released GameScaling, Qwen3.5 VLMs continued-pretrained on 1,600 hours of PC gameplay.
  • Eight Apache-2.0 checkpoints on Hugging Face at 0.8B, 2B, 9B, and 27B parameters.
  • Each call emits 200 ms of actions: 6 key-group steps of 33 ms plus mouse XY and scroll.
  • S1 variant trains on Minecraft only, S3 mixes 27 percent Minecraft with other games for generalization.
  • Ships inference client, vLLM serve script, fixed system prompt template, and a Doom Battle-1 example.
  • No training code, gameplay corpus, or RL loop released; paper covers scaling analysis.

Researchers from miHoYo’s Honkai AI R&D group, Tsinghua University, the University of Chinese Academy of Sciences, and the University of Hong Kong have released GameScaling, a family of vision-language models that observes game frames and produces keyboard and mouse actions. The Apache 2.0 release includes eight checkpoints, an inference client, prompts, and vLLM serving scripts.

Across four model sizes and two training mixtures, the project tests how game-playing performance changes with parameter count and gameplay diversity. The researchers further trained Qwen3.5 vision-language backbones on 1,600 hours of gameplay footage aligned with the keypresses and mouse movements recorded for each frame.

Screen in, input events out

At each inference call, the model receives the current game frame, a short history of earlier frames, and recent actions. It returns the next 200 milliseconds of input: six keyboard states lasting 33 milliseconds each, plus one mouse movement and scroll command for the entire interval. The model therefore chooses an action chunk at roughly 5 Hz while preserving key states aligned with a 30 FPS game loop.

The output follows a compact grammar wrapped in special tokens:

<|action_start|>X Y Z ; g1 ; g2 ; g3 ; g4 ; g5 ; g6<|action_end|>

X and Y specify mouse movement in thousandths of the screen’s width and height. Z specifies scroll notches, clipped to ±5. Each group from g1 through g6 lists as many as four keys held during that 33-millisecond step. Malformed output is parsed as a no-op, allowing the integration layer to avoid executing incomplete commands.

An optional reasoning channel allows the model to emit <think>...</think> before its action. The supplied system prompt instructs it to use reasoning when the game state changes materially and to emit actions directly during routine play.