Kunlun Unveils World Model Matrix

机器之心 · wechat · 2026-07-19

This WAIC feature focuses on Kunlun's "first year of world models," highlighting three main tracks: **World Model Foundation, Embodied AI, and AI Music**. For embodied AI, the article introduces **Riemann-1.0**, defined as a universal robotic brain/WorldActionModel. It unifies action generation, environmental state evolution, and future visual prediction into a causal generative framework. Trained on human egocentric video, exoskeleton/glove data, and robotic trajectories totaling **232,000 hours**, the reported benchmark results include LIBERO **99.0%**, RoboCasa-365 **62.6%**, and RoboTwin2.0 **94.3%**. Adding human egocentric pre-training reportedly boosts long-horizon task success rates significantly. Regarding the world model foundation, the spotlight is on **Matrix-Game3.5**. Designed for long-term consistency, real-time interaction, and open-world generation, it introduces mechanisms like PatchMemory, PRoPE, and decoupled dynamic-static memory. Despite having only **5B parameters**, the model can achieve real-time streaming generation at **720P and roughly 20FPS** on a single GPU, and it has been open-sourced. This world model is currently applied in gaming, digital twins, simulations, and embodied AI training environments. In content generation, the article details two music models, **MurekaV9.5** and **O3**, emphasizing their "restrained arrangement" and "reflection capabilities" to better align with human aesthetics. The article provides several standardized evaluation metrics, such as a **61.0%** good vocal rate, a **97.0%** good prompt control rate, and a **95.7%** full genre representation rate.

Related event: Kunlun Releases and Open-Sources Matrix-Game3.5 World Model(2 posts)→

Original post →

More from Embodied

Embodied channel →