GameScaling releases open models for controlling games
GameScaling releases eight open-weight models that turn visual observations into short bursts of keyboard and mouse controls across diverse virtual worlds.
A research team involving miHoYo’s Honkai AI R&D group and several universities has released GameScaling, a family of open-weight models designed to control games through keyboard and mouse actions. The release includes eight checkpoints ranging from 800 million to 27 billion parameters, alongside inference code and examples for running the system 1.
GameScaling treats a game as an interactive visual environment rather than a sequence of isolated screenshots. Each call receives the current frame, a short history of earlier frames and actions, then predicts the next 200 milliseconds of keyboard and mouse input. That gives the model a compact control loop: look at the world, decide what to do, issue a short burst of input and repeat.
The project’s demonstrations cover several kinds of virtual world. The released material includes examples involving Minecraft, Doom and escape-room-style puzzles, while the project graphic shows agents moving between very different environments, from driving and platforming to strategy and board-game scenes. The team’s stated goal is to study whether continued pre-training across diverse games produces agents that transfer more reliably between worlds.
What is being released
The eight checkpoints are based on Qwen3.5 vision-language models and are split between smaller and larger model sizes. The S1 versions use Minecraft-only training, while S3 mixes gameplay from multiple games. The repository contains the inference client, prompts, tests and examples; the training code and the gameplay dataset are not being released.
The project is also unusually clear about licensing boundaries. The repository says its weights and inference code are available under Apache 2.0, but game footage, titles and other game assets remain the property of their respective publishers. That distinction matters for anyone building demonstrations on top of the release: the model licence does not grant permission to reuse commercial game content.
Our opinion
GameScaling is interesting because it treats generalisation as the central problem rather than presenting one impressive game demo as proof that game-playing AI is solved. A model that can move between worlds with different physics, interfaces and visual languages would be more useful than one trained to repeat a single route. The open weights also make the claim testable, although the largest checkpoints are far beyond ordinary consumer hardware.
The release is best understood as a research platform, not a ready-made autonomous player. Its performance will depend on the games, prompts, hardware and evaluation setup, and the private training data limits how completely outsiders can reproduce the work. Even so, the code, weights and examples give researchers a concrete way to explore embodied agents in games rather than relying on a closed demonstration.