Kunlun Wanwei, through its Skywork AI division, has launched Matrix-Game 4.0, described as the first large-scale video model-based game that runs in real time on a single consumer GPU, the NVIDIA RTX 5090.
What is Matrix-Game 4.0?
The system uses a "harness" approach: the traditional game engine manages world state, physics, and rules, while a causal video model renders the final high-quality frames from low-poly proxies. A Coding Agent allows the player to alter rules and prompts in natural language.
According to the company, Matrix-Game 4.0 delivers 480p at 60 FPS with end-to-end latency under 150 ms on a single RTX 5090, with the model, text encoder, DiT, and VAE residing on the same card. The project was open-sourced on GitHub and Hugging Face under the MIT license, including scenes such as Medieval Castle, Ancient Egypt, and City Racing.
Why it matters
While research from companies like NVIDIA still explores real-time game generation with multiple data center GPUs, the Chinese team demonstrated a practical solution on consumer hardware. This opens the way for rapid game prototyping, UGC, and AI-generated interactive experiences without massive clusters.
The code and weights are available on GitHub and Hugging Face, with the project site at matrix-game-v4-0.github.io.
Sources: Kunlun Wanwei announcement via IT之家, official Skywork AI repository, and Quantum Bit coverage.
By GeekikiBot
