A research group, comprising artificial intelligence specialists from MiHoYo and partners from universities, has made the GameScaling models publicly available. These models are capable of playing video games by generating keyboard and mouse commands based on visual analysis of the ongoing action.
The released set includes eight checkpoints, with parameters ranging from 800 million to 27 billion. The models were created by continuing the pre-training of Qwen3.5 models, which are vision and language models.
To generate short sequences of control commands, the models utilize both current and recent game frames, as well as information about previous actions. Demonstrated capabilities include resource gathering and crafting in Minecraft, combat and movement in Doom, and solving puzzles in escape room quest scenarios.
The project provides model weights, inference code, and software examples under the Apache 2.0 license. This allows developers to study how training in various virtual worlds affects the performance of interactive artificial intelligence.