Physis
Category: Foundation Models / LLMs
Building the next-generation general world foundation model that truly understands real physical laws and predicts next physical states, bridging AI from digital to physical worlds. Physis was founded in 2026. The company is led by Chen Boyuan (陈博远). Based in Beijing, China. Team size: 11-50. Total funding raised: $110.0M. Latest round: Seed. Key investors include Hillhouse Ventures (高瓴创投), Yanyuan Ventures (燕缘创投), Matrix Partners China (经纬创投), Guanghe Capital (光合创投), 5Y Capital (五源资本), BAI Capital, Zhongding Capital (钟鼎资本), Ant Group (蚂蚁集团).
- Founded
- 2026
- Headquarters
- Beijing, China
- Team size
- 11-50
- Total funding
- $110.0M
Value proposition
A general world foundation model that predicts next physical states (P(s'|s,a)) rather than next tokens or pixels, with built-in physical correctness, causal reasoning, and RL-verified physical consistency — serving as the universal cognitive engine for all physical AI applications.
Products and solutions
Physis-v0.1 (悟界·Physis-v0.1) — world's first general world foundation model, released June 2026 at BAAI Conference, supports embodied intelligence, physical simulation, interactive worlds, and scientific prediction, W0-W5 world model capability分级 system defined, flagship model planned for late 2026 release with open-source components
Unique value
First to pursue "Next Physical State Prediction" paradigm with native action injection in latent space + RL-verified physical correctness, rather than relying on video generation or 3D reconstruction approaches. One-for-all foundation model that understands universal physical laws (gravity, collision, fluid dynamics) applicable across embodied AI, industrial simulation, gaming, and scientific prediction.
Target customer
Embodied intelligence/robotics companies; industrial simulation & digital twin platforms; game physics engines; autonomous driving; scientific prediction (materials, fluids, structures); serious industrial applications requiring physically accurate world modeling.
Industries served
Embodied AI/Robotics, Industrial Simulation, Gaming/Interactive Worlds, Autonomous Driving, Scientific Research (materials, fluids, structures)
Technology advantage
Proprietary architecture with native action injection in physical latent space (not pixel space); RLVR (Reinforcement Learning with Verifiable Reward) for physical consistency; data pyramid (L1-L3) for efficient physical interaction data; self-built data production pipeline for rare physical edge cases; demonstrated scaling laws in physical state prediction; team of 10+ international Olympiad gold medalists and 50+ top conference papers
How they differentiate
Unlike World Labs (3D spatial reconstruction) and AMI Labs (JEPA representation learning), Physis builds world models from first principles using RL + native action injection in latent space, with physical verifiability as the core metric rather than visual fidelity. They prioritize physical correctness over visual quality, using RLVR to ensure models learn causal physical laws rather than statistical correlations.
Main competitors
World Labs (Fei-Fei Li, spatial intelligence/3D world models), AMI Labs (Yann LeCun, JEPA architecture for world models)
Key partnerships
BAAI (Beijing Academy of AI / 智源研究院) — co-developed and released Physis-v0.1 at 2026 BAAI Conference, NVIDIA — co-presented at World Model Forum alongside AMI Labs, Ant Group — strategic investor, Peking University — talent pipeline and joint lab planned via Yanyuan Ventures
Notable customers
Not publicly disclosed (company is in early stage, pre-commercialization)
Major milestones
2026-02-14: Company founded, 2026-03-26: $10M+ Angel round led by Hillhouse Ventures & Yanyuan Ventures, 2026-06-12: Released Physis-v0.1 at 8th BAAI Conference, billed as world's first general world foundation model, 2026-06-17: $100M+ Seed++ round closed with Matrix Partners, Guanghe Capital, Ant Group and others
Growth metrics
Founded Feb 2026; raised ~$110M+ in two rounds within 4 months; team of ~30 with 10+ Olympiad gold medals and 50+ top conference papers; Physis-v0.1 model released June 2026
Market positioning
Early-stage pioneer in the emerging "world foundation model" category, positioning as the physical-world equivalent of large language models. Competing against academic labs (World Labs, AMI Labs) and potentially big tech, with a unique RL + world model approach that differentiates from both the 3D reconstruction and video generation paths.
Geographic focus
China (Beijing-based, strong PKU ties, BAAI collaboration); global ambitions with English-language research output and international partnerships
About Chen Boyuan (陈博远)
Peking University Yuanpei College AI undergraduate (class of 2026); NeurIPS Oral & Spotlight, ACL Best Paper Award winner; 2,300+ Google Scholar citations; PKU May Fourth Medal & Person of the Year; Lead of BAAI Behavioral World Model Innovation Center
Official website: https://cby-pku.github.io/