Tsinghua University, in collaboration with miHoYo and other institutions, has jointly open-sourced GameScaling, the first general-purpose game model series. The release covers eight models across four size tiers—0.8B, 2B, 9B, and 27B—with weights and inference code made available under the Apache 2.0 license. These models can generate keyboard and mouse inputs based on game visuals, task prompts, and interaction history, and they support interactive tasks across multiple games with distinct mechanics. Some small- and medium-scale models can achieve real-time game interaction on a single consumer-grade or professional graphics card, and they perform excellently on Minecraft-related benchmarks. Through controlled-variable experiments, the team systematically investigated the validity of Scaling Laws in scenarios with vastly different game mechanics. They found that training proceeds in two stages: an interface stage for adapting to action formats, and a world stage for learning world knowledge. Increasing data diversity can significantly enhance cross-game transferability, and larger models absorb experience more efficiently; however, long-term training on a single game may instead lead to performance regression in unfamiliar worlds with large distribution gaps. Existing Scaling Laws can effectively estimate learning gains in worlds covered by training, but for unseen worlds with large differences, stable improvements cannot be achieved solely by scaling up model and data size. In real-time game interaction evaluations, the model outperforms multiple mainstream frontier large models, validating the importance of low-latency decision-making for game intelligence and providing a testable research starting point for subsequent work on general game intelligence.