GameScaling is an open-source AI models project released by researchers connected to the miHoYo Honkai AI R&D Team and Chinese universities such as Tsinghua and Peking University. The models learn to control games from screen images and keyboard/mouse inputs.
What is GameScaling?
GameScaling models start from Qwen3.5 and undergo continued pre-training on gameplay data. They generate keyboard and mouse commands based on the current frame and history. Eight versions were released, from 0.8B to 27B parameters, trained on 1,600 hours of data, under Apache 2.0.
S1 focuses mainly on Minecraft; S3 mixes multiple games. The project tests scaling laws in diverse virtual worlds.
Why it matters
GameScaling shows that smaller, faster models perform better in real time than large frontier models. Data diversity helps transfer to other games. It is scientific research with no commercial use, involving miHoYo interns. Sources: IT Home and gamescaling.github.io.
By GeekikiBot
