Skywork AI, the AI unit of Kunlun Tech, released Matrix-Game 4.0 on October 15, the latest version of its open-source interactive world model series. Inference code, a browser demo and the source code of three playable games are on GitHub, with model weights and game assets on Hugging Face. The new version is licensed under Apache-2.0.
The release changes how the series approaches game worlds. Instead of asking a video model to remember the world and keep track of the rules while it generates frames, Matrix-Game 4.0 keeps gameplay, physics and player input inside an executable game engine. The engine outputs a rough low-poly view of the scene, and a causal, few-step video model turns that view into detailed visuals in a streaming fashion. The team describes the approach as putting a harness on the video model: code decides what happens, and the model decides what it looks like.
That split means state such as health, item ownership and mission progress is recorded explicitly in code, so an occasional visual glitch does not rewrite the game. In a demo reported by QbitAI, a picked-up shield absorbs damage before the player's health drops, and the game ends when health reaches zero. Players can also ask a coding agent in natural language to change the game code, for example raising a weapon's fire rate or turning a time trial into a pursuit.
The first open-source scenes are Medieval Castle and Ancient Egypt, both third-person shooters, and City Racing, a driving game. To train the renderer, the team built paired data in Unreal Engine 5, matching structural frames with fully detailed ones, and feeds the model dynamic captions describing visible objects and their state. The video model is adapted from Wan2.2 5B Fun Control and also draws on spatial memory reference images and cached recent history.
Skywork calls the system the first game world model that runs in real time on a single consumer graphics card. According to QbitAI, on one RTX 5090 it reaches about 83 frames per second with one denoising step and about 57 with two, runs games at 60 frames per second, responds to a key press in under 0.15 seconds and peaks at about 18.5GiB of video memory. Those figures come from the company. The GitHub documentation ships a dedicated RTX 5090 preset using FP8 kernels and CUDA graphs, and its own measured table lists roughly 38 to 39 frames per second of model processing in online sessions for the shooter scenes. Output resolution in the demos is 480p.
Earlier versions added action control in 1.0, real-time generation in 2.0 and long-horizon memory in 3.0, released in March.