miHoYo and Top Universities Open-Source GameScaling Vision-Language Models
miHoYo's Honkai AI research team, alongside several top universities, has released and open-sourced the GameScaling vision-language model series, which ranges from 0.8B to 27B parameters and is based on continued pre-training of Qwen3.5. These models are designed to ingest game footage, history, and actions to generate keyboard and mouse controls for navigating virtual worlds like Minecraft and Doom. This release provides a valuable open-source resource for studying scaling laws in embodied AI agents operating within dynamic game environments. By releasing models across various sizes with support for standard tools like vLLM under an Apache 2.0 license, it lowers the barrier for researchers exploring real-time game interaction and advanced AI decision-making. The model line-up consists of 8 checkpoints across 0.8B, 2B, 9B, and 27B parameter versions, split into single-game (Minecraft-only) and multi-game variants trained on 1,600 hours of gameplay data. Each model call generates a 200-millisecond sequence consisting of six 33-millisecond action steps alongside mouse movement and scroll instructions.
## BACKGROUND
Vision-language models (VLMs) are artificial intelligence models capable of understanding and processing both visual inputs, such as images or video frames, and textual information. Scaling laws in AI describe the mathematical relationship between model performance and key factors like parameter scale, dataset size, and compute resources.