miHoYo, Tsinghua University, and others open-source GameScaling, a game control model
miHoYo's Honkai AI team, collaborating with Tsinghua University and others, has released GameScaling. Building on Qwen3.5's vision-language model through continued pre-training for keyboard and mouse controls, the models (0.8B–27B) output key/mouse actions for the next 200 milliseconds. The paper evaluates data volume, diversity, and model scale across 120 training configurations. Eight checkpoints and an inference client are open-sourced under Apache-2.0, excluding training code.
miHoYo, in collaboration with Tsinghua University and other institutions, has open-sourced GameScaling, a game control model.
The miHoYo Honkai AI team, together with Tsinghua University and others, has released GameScaling. It is based on the Qwen3.5 vision-language model and undergoes continued pre-training for keyboard and mouse controls. The model comes in sizes ranging from 0.8B to 27B, with each output predicting keyboard and mouse actions for the next 200 milliseconds.
The paper investigates the effects of data volume, diversity, and model scale using 120 training configurations. The findings show that more data improves performance within the same source domain, while diversity aids transfer to similar games but has limited impact on worlds with significant differences. Eight checkpoints and an inference client are open-sourced under the Apache-2.0 license, though the training code is not included.
