Skywork AI Details Matrix-Game 3.5 World Model

Kunlun Tech chairman and CEO Fang Han said at WAIC on July 19 that 2026 is the “first year of world models,” and its subsidiary Skywork AI detailed Matrix-Game 3.5 in a press release. The interactive world model is designed to move from generated clips toward systems that track objects and spaces across user actions.

Matrix-Game 3.5 uses a Patch Memory mechanism that stores spatial patches with three dimensional coordinates instead of storing full frames. Skywork AI said the model also uses PRoPE geometric position encoding and Warped RoPE to track position in a world coordinate system.

The company said the 5B parameter model can generate streaming video at 720P and about 20 FPS on a single GPU, with one minute memory. Its training pipeline includes more than 5 million video segments, over 10,000 hours of training data, and more than 1,200 game scenes with camera pose, depth maps, and text descriptions.

Skywork AI said most of the core interactive functions were added without increasing the base model parameter load, apart from a character Reference Adapter. The company described the design as modular, with components for geometric encoding, memory, camera control, and dynamic static separation.