In the World Model Department, the Space Intelligence Generative World Model represented by Li Feifei has been acquired by AMD for $8.5 billion, while the other faction is the Representational World Model, which has numerous companies and can be researched up to 15. The representational world model does not "draw the next frame" in pixel space, but encodes the image into an abstract semantic vector, predicts future states in latent space, and directly serves decision-making and planning. Contrary to the generative approach of Sora, which does not reconstruct pixels and does not consider high-dimensional images, it is more efficient and less prone to being biased by textures and lighting. Two lineages, JEPA (Meta FAIR): encoder enters the latent space, predictor only predicts future representations, self supervised, and does not reconstruct pixels. The lineage is I-JEPA → V-JEPA → V-JEPA 2 (open source in 2025.6, masking 75% of videos and predicting abstract features of missing parts, training efficiency is about 1.5-6 times that of traditional models, supporting zero sample planning) → AdaJEPA (2026.7, adaptive when testable) → LeWorldModel Dreamer ：， Plan based on imagination. DreamerV3 is the first algorithm to mine diamonds from scratch in Minecraft without relying on human data. Dreamer (DeepMind): In the field of subspace, st+1=f (st, at) s_ {t+1}=f (s_t, a_t) s_ {t+1}=f (s_t, a_t) At home and abroad, 15 early-stage companies are laying out their strategies in this representation: 15 layouts: 1/Meta FAIR: The birthplace of JEPA, V-JEPA 2 is open source. 2/AMI Labs (Paris, January 2026.1): JEPA industrialization, targeting industries such as robotics and healthcare, with research before products. 3/Brain Rock (Shanghai, 2025.6): Zhu Senhua, cognitive world model CWM, self proclaimed JEPA enhanced version, with a chain of perception → cognition → prediction → planning → action → learning, prioritizing Japanese industry and "one brain, multiple machines". 4/Coronal Opening (Beijing, March 2026.3): Background of Tsinghua Intelligent Driving, LaMPA， The triple representation of environment/ontology/experience+block diffusion+world reward model, with a seed round totaling hundreds of millions of yuan, has been assembled with high precision on the server. 5/Zhicheng AI (Hangzhou): self-developed world model+ontology (TR4/TR5/TR6/XR1), integrating software and hardware. 6/Vision Future (Shenzhen, 2025.8, IDEA incubation): Zhang Lei, nearly 100 million yuan angel wheel, hidden space world model with "object action" as the core, product EgoWin. 7/DeepMind Dreamer: The academic origin of the latent space world model; Genie 3 is biased towards generation/interaction, and Dreamer is the representative representation. 8/Unbounded Power (Beijing, 2025): MWA, Hidden Space+Reinforcement Learning, accumulated hundreds of millions of dollars in financing in the first half of the year, RoboCasa ranked first in 75.2% globally, with orders of nearly 100 million dollars. 9/Super Brain Future: WorldDreamer V4 (July 27, 2026), sharing the same world model with multiple agents, "One World One Model", RoboCasa and WorldScore blind test first place. 10/Qianjue Technology (Beijing, 2023, Tsinghua Brain Center): Predicting low dimensional physical trajectories rather than images, only making "robot brains", claiming to have access to 100000 terminals and a 2026.6 billion yuan A-round. 11/Inverse Matrix (Beijing, February 2026.2): Wu Jie · Physis, active intervention in hidden space physics causality, seed++round exceeding billions of dollars, team almost all born in 1995/00. 12/LiberAI (Beijing, May 12, 2025): Liu Songming, aligned forces, positions, and trajectories into the video representation space, added Newtonian induction bias, and raised hundreds of millions of yuan in cumulative financing. 13/Stardust Intelligence (Shenzhen): Lumo-2 (July 15, 2026), anticipates the world before taking action, a single consumer grade graphics card can achieve real-time closed-loop, and 22 household tasks can be automated on the real machine. 14/It is Shi Zhihang (Shanghai): AWE 3.5, WAIC 2026 SAIL Star, with smooth hidden space movements that reduce shaking by over 45%. It is explicitly required to predict the world state after the action and has entered the automotive wiring harness production line. 15/Cross dimensional Intelligence (Shenzhen, 2021): Jia Kui, DexWorldModel predicts future features in DINO semantic space, B-round 1 billion yuan, valuation exceeds 10 billion yuan, preparing for Hong Kong IPO; WorldArena Track 2 is ranked first globally.