miHoYo and Tsinghua open GameScaling: eight Apache-2.0 game-playing models (0.8B–27B) that output keyboard-and-mouse actions; 27B tops Minecraft specialists and 9B beats frontier models in real-time play

Why it matters

【Key points】

On Oct 22, miHoYo's Honkai AI team, working with Tsinghua, UCAS, HKU and Peking University, opened GameScaling: eight game-playing models fine-tuned from Qwen3.5 at 0.8B, 2B, 9B and 27B. Weights and inference code are Apache 2.0. Each model reads game frames and outputs the next 200 ms of keyboard and mouse actions, and can reason first. S1 variants trained on Minecraft only; S3 variants on a multi-game mix that is 27% Minecraft. Each used 1,600 hours of gameplay.

Across 120 training runs the team fitted a scaling law over data volume, diversity and model size. On Minecraft's MCU benchmark, 27B-S1 averages 63.3% vs 50.7% for the best Minecraft specialist (GameWAM). In real-time play, 9B-S3 averages 25.3 vs 19.2 for Gemini 3.8 Flash; frontier models take 4–24 s per step, GameScaling 0.13–0.69 s. Training data and training code are not released.

【Verifiable】

https://huggingface.co/THUSI-Lab/GameScaling

https://gamescaling.github.io/

https://arxiv.org/abs/2610.28698

https://github.com/THUSI-Lab/GameScaling

https://www.ithome.com/1/016/575.htm

【Our read】

1. It sits between Minecraft-only specialists (VPT, JARVIS-VLA) and frontier-model agents (SIMA, the Gemini 4 Argon Minecraft run we covered Oct 16): small open models, continued pretraining, released weights.

2. When the game pauses for the model, frontier models still lead, and they double its score on distant 2D games. In real time they lose about 90% on ViZDoom and MCU. For NPCs and robots, speed may matter more than size.

3. Caveats: GameScaling ran on the team's own H200s while frontier models were called over public APIs, all scores are self-reported, and training can't be reproduced without the data.