At WAIC 2026, the world model was everywhere, and Kunlun Wanwei arrived with a claim that China, not just Google, now sets the pace.
Its Matrix-Game 3.5 targets the embarrassment that humbled the field: object permanence. On a benchmark called MemoBench, where a hidden object must reappear in the right state, ten leading models scored a maximum of 0.58 out of 1. A three-year-old child scores full marks.
Remembering space, not frames
Kunlun’s fixes are three. First, it swaps time-indexed memory for a geometrically aligned, patch-level 3D memory, so the model recalls ‘an object at coordinates x, y, z’ rather than ‘something in frame five’. Second, it encodes the camera projection matrix into position embeddings, so the model knows which viewpoint a pixel comes from. Third, it decouples static scene from moving subject, and bakes causal inference into the training objective by generating state and action jointly.
The result is a 5-billion-parameter model that runs at 720p and roughly 20 frames per second on a single GPU, fast enough that intervention becomes real-time control rather than post-hoc video.
An open challenge to Genie
Kunlun has released models in this series from the start, positioning Matrix-Game as the domestic answer to Google’s Genie. The validation is unusual: NYU assistant professor Saining Xie, a co-author of the JEPA world-model line with Yann LeCun, built the multi-agent video world model Solaris on Matrix-Game 2.0’s open base, and NVIDIA with Zhejiang University published work on Matrix-Game 3.0.
The ambition reaches past games. A robot can practise dropping ten thousand bowls inside the engine for the cost of tokens; an autonomous car can face rain, night and a sudden jaywalker on demand. Traditional game engines die on contact with physics; Kunlun wants its model to make the world feel alive.
Matrix-Game 3.5 is not the finish line. It is a lever, prying world models out of video generation and into physical control.
Translated and adapted from Leiphone (leiphone.com).