miHoYo researchers open-source GameScaling, with eight models for keyboard-and-mouse gameplay
Researchers from miHoYo’s Honkai AI team and university collaborators have released GameScaling, a set of vision-language models that use game images and recent actions to generate keyboard-and-mouse controls. TechNode reports that the eight Qwen3.5-based checkpoints range from 800 million to 27 billion parameters and are released with inference code under the Apache 2.0 licence; the project does not include training code, according to its GitHub repository.
Bottom line — GameScaling’s eight checkpoints are downloadable, but its training code and gameplay data are not.
Go deeper 6
-
Each decision covers 200 milliseconds of play, split into six control steps of 33 milliseconds, according to the project documentation.
-
TechNode reports demonstrations including resource gathering and crafting in Minecraft, and movement and combat in Doom.
-
The project repository says the models use current and recent game frames alongside previous actions.
-
According to the project repository, S1 checkpoints train on Minecraft alone, while S3 uses a mixture of games.
-
The 36Kr article says the paper is an arXiv preprint and has not yet undergone peer review.
-
In a separate account of the paper’s results, 36Kr reports that varied training improved performance on similar 3D games, while Minecraft-only training could weaken performance on other benchmarks.
www.amazon.com