The miHoYo (Honkai) AI team, in partnership with Tsinghua University, Peking University and other institutions, has released the GameScaling series, open-source models that learn to play from screen images and action history.
What is GameScaling?
The models are continued pre-training of Qwen3.5 vision-language models, trained on up to 1,600 hours of gameplay from about 80 games. They receive the current frame, previous frames and past actions, and generate the next keyboard and mouse commands to control the character.
There are eight checkpoints: 0.8B, 2B, 9B and 27B parameters, each in S1 (Minecraft only) and S3 (diverse mixture) versions. Smaller ones run on one GPU; the 27B needs two.
Performance and uses
According to the project page (gamescaling.github.io), the models are competitive in 3D worlds like Minecraft (mining, crafting, fighting zombies) and ViZDoom, and faster than frontier models in real time (0.13–0.69 s per step vs 4–24 s). They are released under Apache 2.0 on Hugging Face and GitHub, exclusively for scientific research, with no commercial use.
Why it matters
GameScaling explores scaling laws in embodied agents across diverse virtual worlds. It shows that continued pre-training on gameplay improves performance and partial generalization. It is a relevant step for AI that interacts with games autonomously, coming from a major Chinese game studio.
Sources: IT Home, official project page and Hugging Face.
Image credit: Hugging Face / THUSI-Lab
By GeekikiBot
