Skywork AI, the AI unit of Kunlun Tech (China), launched Matrix-Game 4.0 on October 15, 2026. The company describes it as the first video model game capable of real-time play on a single consumer GPU.
What happened?
The system splits responsibilities: a traditional game engine handles physics, rules, world state and player input, outputting a simple low-poly view. A causal video model (related to Alibaba’s Wan architecture, roughly 5B parameters) turns those structural frames into detailed images in a streaming fashion.
According to Skywork, IT Home and QbitAI, on an RTX 5090 the system reaches about 60 FPS at 480p, with end-to-end latency under 150 ms (around 112 ms in measurements) and peak VRAM around 18.5 GiB. A coding agent lets users change gameplay rules in natural language without retraining the model.
The project is open-sourced on GitHub and Hugging Face under Apache-2.0, with inference code, a browser demo and three playable scenes: Medieval Castle, Ancient Egypt and City Racing.
Why it matters
Most previous interactive world models tried to make the video model maintain both state and rules, which required data-center hardware. By letting the game engine act as the "referee" and the model only as the "painter", Matrix-Game 4.0 drastically lowers the compute cost and makes generative game prototypes more practical for individual developers.
Sources: Skywork AI official announcement, IT Home, QbitAI and the GitHub repository SkyworkAI/Matrix-Game.
Image credit: Skywork AI / coverage reproduction
By GeekikiBot
