ModelsRobotics 🇨🇳 27.07.2026 07:02

Kunlun Wanwei Bets on Two World Model Paths: Real-Time Interactive Content and Robot Pre-Simulation

Kunlun Wanwei released Matrix-3.5, Mureka V9.5/O3, and Riemann-1.0, targeting interactive content, music, and robotics. CEO Fang Han calls 2026 the 'first year of world models,' envisioning two trajectories: real-time content generation and pre-simulation for robots. The models aim to understand environments, remember past events, and predict action outcomes.
Kunlun Wanwei simultaneously released three models: Matrix-3.5 for games and video, Mureka V9.5 and O3 for music, and Riemann-1.0 for robotics. CEO Fang Han stated that world models are entering industry along two paths: one in virtual worlds where games and videos will be transformed by real-time interactive generation, and another in the physical world where robots must internally simulate environmental changes before acting. Matrix-3.5 upgrades long-term memory, spatial consistency, and inference efficiency using patch memory based on 3D spatial information, separating dynamic and static elements. Mureka V9.5 focuses on reducing the 'model-like' feel for more authentic music, while Mureka O3 introduces reasoning and reflection mechanisms. Riemann-1.0 jointly generates actions and future visual predictions in a unified diffusion Transformer architecture, trained on over 200,000 hours of human and robot data, achieving 80-90% success rates in real-world tasks. The company also discussed AI's impact on content creation, noting that AI reduces costs but requires creator judgment, and that world models face challenges in long-term decision-making and data acquisition.
Сокращения
VLA = Vision-Language-Action Model
Source: InfoQ 中国 — original
Our earlier posts on this topic ↓
Fresh news