miHoYo's Honkai AI research team, working with Tsinghua University, the University of Chinese Academy of Sciences, the University of Hong Kong and other institutions, has released GameScaling, a family of game-agent models that output keyboard and mouse actions directly. The paper went up on arXiv on October 20, and the model weights and an inference client are now public on Hugging Face and GitHub under the Apache-2.0 license.
The models are built on Alibaba's Qwen3.5 vision-language model and further pretrained on gameplay video that comes with keyboard and mouse logs. Show it a few frames and it returns which keys to press next and where the mouse should move.
How actions are encoded
On each call, the model outputs the next 200 milliseconds of actions, split into six 33-millisecond steps, plus the mouse movement and scroll wheel input for that whole block of time. Mouse displacement is expressed in thousandths and capped at plus or minus 1,000; the scroll wheel is capped at plus or minus 5. Each step can press up to four keys at once, and the key vocabulary used in training contains 74 keys.
The model can either think before acting or output actions directly, switched by prefilling a <think> tag. The project recommends deploying with vLLM, with up to eight images per request; the default is five historical frames plus one current frame.
The models come in four sizes: 0.8B, 2B, 9B and 27B. The first three run on a single GPU, while the 27B model's bfloat16 weights come to about 54.7 GB and need two GPUs with tensor parallelism. Each size has an S1 and an S3 version. S1 uses Minecraft data only; S3 is a balanced mix in which Minecraft accounts for 27% and the rest comes from other games. That makes eight checkpoints in total.
What 120 configurations revealed
The core of the paper is a set of scaling experiments. The authors designed 120 training configurations across three axes: data volume, diversity and model size. Data budgets range from 50 to 1,600 hours in six tiers, and there are five data-mixing recipes.
The findings come in several layers. More data improves performance within the same game. Diversity helps the model transfer to similar games, but the gains are limited when the target world is very different. When training on a single game, a larger model makes better use of data, yet generalizing to distant games becomes harder.
Training also splits into two phases. In the early "interface phase," the model learns how to operate a keyboard and mouse, and those gains carry across games. In the later "world phase," the gains narrow to the same game and close relatives. On that basis the authors fitted empirical scaling laws relating validation loss to the three variables, and they say the models can hold their own against Minecraft-specific models and frontier models in 3D worlds, performing best in real-time settings.
The paper's conclusion is blunt: imitation learning has a ceiling, gaps remain on distant worlds, and continued pretraining alone is not enough to build a general embodied agent.
What developers in China can use
The release offers a few ready-made conveniences for Chinese teams. The base is Qwen3.5, whose tooling is familiar across domestic clouds and local deployments, and under Apache-2.0 the weights and inference code can be modified and integrated.
The limits are spelled out as well. The repository contains no training code, the training corpus is not public, and no game footage is released; game copyrights belong to their publishers, and the license grants no rights to any game content. The project states it is for research only and not for commercial activity. The paper itself is posted on arXiv under CC BY-NC-SA 4.0, a separate license from the model's Apache-2.0, so teams planning commercial use need to review the two separately.
In practice, teams working on automated game testing, companion bots or embodied-agent research can try the 0.8B or 2B version on a single GPU and see how far it gets in their own game. Anyone who wants to keep training on their own data will have to build the training pipeline themselves.
The README does not list the full set of supported games. The only examples are Minecraft and a Doom level in ViZDoom, and neither the paper nor the repository says how the models perform in miHoYo's own games.
Sources: arXiv paper page, CocoLoop, GameScaling GitHub repository; the repository README was used to verify the checkpoint count, weight size and action encoding format.