Manifold AI’s WorldScape Tops WorldScore, Outperforming Li Fei‑Fei’s Team

WorldScore Benchmark Victory – The international WorldScore leaderboard, recognized as the ultimate test for general world‑model bases, recently updated its rankings. Manifold AI’s self‑developed WorldScape model surged to the global first place, surpassing competitors including Li Fei‑Fei’s team, MIT, Alibaba, Runway, Zhipu, MiniMax and Tencent.

Parameter Efficiency and Spatial Intelligence – Despite having a model‑parameter scale an order of magnitude smaller than the other top‑ranked models, WorldScape achieved the highest spatial‑intelligence density worldwide, demonstrating that a compact model can deliver superior performance.

Addressing the Core Challenge of World Models – Traditional world models excel at visual fidelity but falter on complex physical dynamics and multi‑step control, often producing noticeable artifacts. WorldScape tackles the hardest physical and interaction metrics, establishing a clear advantage that enabled its leaderboard dominance.

Deep Fusion of Generation and Control – The model adopts a unified action‑world‑state modeling framework that merges spatial displacement and object interaction into a single generation process. This eliminates inconsistencies caused by stitching multiple modules and simultaneously supports spatial navigation and object manipulation.

Stable 3‑D Structure – During training, WorldScape explicitly incorporates 3‑D geometric awareness and constraints, ensuring that generated results maintain consistent spatial structure across continuous interactions. This design effectively mitigates the geometric drift and structural collapse that commonly occur in long‑duration generation.

Real‑Time High‑Quality Generation – Rather than relying on model compression or resolution reduction, WorldScape uses structured generation and efficient training strategies to achieve near‑real‑time interactive generation (6–16 FPS) on a single GPU, while ranking among the top in visual quality and motion smoothness.

World‑State Memory – WorldScape introduces a geometry‑aware world‑state memory mechanism, allowing the model to share and update spatial information across different timesteps, thereby providing long‑term consistency that distinguishes true world models from mere video generators.

Community Evaluation and Challenges – Manifold AI’s founding team, together with more than ten domestic and international universities, proposed the unified embodied world‑model evaluation suite WorldArena and organized the CVPR 2026 WorldArena Challenge, driving the field from “visual realism” toward “functional usability”.

Product Releases – In early 2024, Manifold AI released WorldScape v0.1, the first world‑model supporting both mobile and manipulation interactions in real time, and positioned it as a pre‑training foundation for robots. The same period saw the launch of WorldScape Policy and the VLA model, which surpasses existing VLA baselines and exhibits few‑shot/zero‑shot execution capabilities.

Data‑Driven Infrastructure – The company built a closed‑loop data pipeline (Ego‑Centric → UMI → RL) that continuously generates high‑quality data through collection, filtering, augmentation and annotation, producing over 100 000 video clips per day. This massive data flow underpins the rapid improvement of model capabilities.

Funding and Team Background – Within ten months of its 2025 founding, Manifold AI secured five financing rounds (including Pre‑A+ led by Shunxi Fund) and amassed a team of former senior engineers from SenseTime, Momenta, Xiaopeng, and other leading AI firms. Founder‑CEO Dr. Wu Wei, a former senior executive at SenseTime, previously led the company’s world‑model research and achieved two consecutive victories over the Waymo SimAgents competition.