IT Home reports that miHoYo's Honkai AI team, together with Tsinghua University, Peking University, the University of Chinese Academy of Sciences, the University of Hong Kong and the National University of Singapore, has released the open-source GameScaling series of game-playing models. The team published eight checkpoints at 0.8B, 2B, 9B and 27B parameters, each in an S1 version trained only on Minecraft data and an S3 version trained on mixed-game data with Minecraft at 27%, all using 1,600 hours of gameplay. According to the project's GitHub description, the models are continually pre-trained from the Qwen3.5 vision-language model and take current and earlier frames plus input history to emit keyboard and mouse actions, with each call covering about 200 ms including six 33 ms steps; tasks span mining, crafting and combat in Minecraft, fighting and pathfinding in Doom, and escape-room puzzles. Code and weights are under Apache 2.0, vLLM serving is supported, and the 0.8B, 2B and 9B weights fit on one GPU while 27B needs two GPUs with tensor parallelism, the report says.