Skywork AI, a unit of the Chinese company Kunlun Tech, launched Matrix-Game 4.0 on October 15, 2026. The system is presented as the first video game model that runs in real time on a single consumer GPU, reaching 480p at 60 FPS with latency below 150 ms on an RTX 5090.
How Matrix-Game 4.0 works
Unlike previous world models that try to manage rules and state only in the neural network, Matrix-Game 4.0 uses a "Harness" architecture. A traditional game engine maintains physics, rules and world state (health, items, progression). The causal video model, based on about 5 billion parameters derived from Wan, renders the low-poly frames into detailed images in a streaming fashion.
A Coding Agent allows changing game rules with natural language. Available scenes include Medieval Castle and Ancient Egypt (third-person shooters) and City Racing.
Performance and availability
According to company data, peak VRAM is about 18.5 GiB, with generation up to 83 FPS in one denoising step. The project is open source under Apache-2.0, with code, weights and assets on GitHub and Hugging Face. Optimized support for RTX 5080 and 5070 Ti was also added.
Why it matters
The approach separates game logic from neural rendering, reducing state consistency issues common in pure world models. This opens the way for more practical prototypes and UGC on consumer hardware. Details on the GitHub repository and the project page.
By GeekikiBot
