Kunlun Wanwei bets on two trajectories for world model deployment: real-time interactive content generation and 'pre-simulation' for robots

Chinese company Kunlun Tech (Kunlun Wanwei) has unveiled three new models operating under the paradigm of world models. Matrix-3.5 is designed for real-time generation of game and video scenes with long-term memory and spatial coherence. Mureka V9.5 and O3 improve the quality of AI-generated music by reducing "robotic" sound and adding emotional variations. Riemann-1.0 is an embodied world model for robots that simultaneously generates actions and predicts environmental changes. The model was trained on over 200,000 hours of data, including first-person video and hand manipulation. According to CEO Fang Han, world models will develop along two paths: in the virtual world (games, video, music) and the physical world (robots that first simulate actions in the model and then execute them in reality). Matrix-3.5 uses a memory mechanism based on spatial patches (Patch Memory) to maintain scene consistency during long interactions. In music models, emphasis is placed on the subjective perception of naturalness of sound. Riemann-1.0 combines action generation and video prediction in a unified architecture, achieving 80-90% success rates in real-world tasks. Experts note that world models will not replace humans in creative work but will change production processes by reducing costs. Technical challenges remain in long-term planning, data scarcity, and simulation reliability for robots.