Ex-Huawei AI leaders found Xirang Kaiwu, raise hundreds of millions in funding at $500M valuation

As digital intelligence reaches saturation, physical world intelligence remains scarce. Two questions persistently raised by investors and industry insiders have emerged: While LLMs (Large Language Models) ushered in the AI 1.0 era of digital intelligence, does the AI 2.0 era of physical intelligence require a new set of foundational models?

How can large models scale into the physical world—enabling robots to enter factories, accelerating drug discovery, improving weather forecasting, and more? What core capabilities are needed?

Xirang Kaiwu has a clear answer: Instead of building robot hardware or point applications, the company is developing a proprietary Large Physics Model (LPM)—a foundational model for the physical world that understands physical laws, predicts state evolution, plans long-horizon actions, and supports closed-loop execution.

Funding and Valuation

Xirang Kaiwu recently announced the completion of multiple rounds of seed and angel financing totaling hundreds of millions of RMB. The round was led by Dunhong Asset, with participation from Huakong Fund, Sanhua Holdings, Ginkgo Valley Capital, Zeran Capital, Benjian Fund, Shanghai Angel Association, and Biaopu Investment. The company is currently valued at $500 million. Proceeds will be used for large-scale pre-training of the native physical AI model, building real-world physical interaction datasets, recruiting core R&D talent, and scenario validation.

A Dual-Engine Team: Industry Execution Meets Theoretical Foundation

Xirang Kaiwu was co-founded by two key figures from Huawei's large model ecosystem:

Li Yin, CEO and founder, graduated from Tsinghua University and served as CTO of Huawei Cloud's large model division. She was a core member of Huawei's large model initiative from its inception and led its industrial deployment. She managed a 400-person team that trained and launched a video large model in just two months using millions of hours of video data and a 10,000-card computing cluster. She also led the deployment of embodied intelligence innovation centers across multiple Chinese cities and delivered large model solutions for autonomous driving, mining, finance, healthcare, and meteorology—generating hundreds of millions in annual revenue.

Professor Zhang Hanwang, currently Huawei's Chief Scientist for Multimodal AI and a Nanyang Presidential Chair Professor at Nanyang Technological University, will serve as co-founder and Chief Scientist. He has long specialized in causal machine learning and multimodal intelligence—the former aiming to push AI beyond "correlation" to "causation," a theoretical cornerstone for machines to truly understand physical world laws. He ranks as the top Chinese scholar globally in causal research on Google Scholar and is among the youngest on the global leaderboard. He was also selected by IEEE as one of the "Top 10 AI Scholars to Watch."

One founder comes from industry; the other from theory. Their combined core competency can be summarized as: possessing both foundational model technical capabilities and the ability to commercialize in industry.

Rebuilding a Physical Foundation Model with the Bellman Equation

In the AI 1.0 era, LLMs liberated humans at the reasoning layer: language, programming, symbols, and tools were all integrated into a single model. But LLMs have a structural limitation—their abstraction of the physical world remains heavily dependent on human intelligence. Without Newton, the model cannot discover F=ma on its own.

AI 2.0 must fill this gap: enabling AI to learn the laws of the physical world. The challenge of physical world decision-making lies in long-horizon tasks—multi-step sequential actions explode combinatorially, making exhaustive search infeasible.

Xirang Kaiwu returns to first principles, grounding its model in the Bellman equation's optimal solution: recursively decomposing long-horizon decisions so that the model learns both the world transition function P and the policy π under a unified objective of maximizing long-term cumulative reward. Architecturally, LPM decouples world simulation from video rendering: a unified autoregressive architecture handles causal history and predicts "what happens," while a diffusion model renders "what it looks like."

LPM = Optimal Solution to the Bellman Equation: enabling the model to approximate decisions that maximize long-term cumulative reward.

Unlike the mainstream "VLM + action head" fine-tuning approach, Xirang Kaiwu pursues native physical pre-training: starting from unified encoding of physical interaction data, state-action-outcome modeling, and long-horizon consistency learning. This allows the model to internalize physical world laws during pre-training, enabling capabilities to scale with model size.

Four Core Strengths

1. Full-stack team with proprietary model development: Core members come from Huawei's large model ecosystem, with end-to-end experience in 10,000-card training, video large models, multimodal AI, and complex industry delivery.

1. First-principles theoretical foundation: Modeling physical decision-making at its essence—using the Bellman equation's optimal solution as the theoretical bedrock, with recursive decomposition to overcome combinatorial explosion in long-horizon decisions.

1. 10,000-card engineering capability: Proven experience in stable training on 10,000-card clusters, trillion-parameter long-duration training, and inference optimization, with cluster availability reaching 99.7%. Hands-on experience tuning the first 10,000-card supernode cluster on Ascend hardware.

1. Mature B2B industrial deployment: Delivered 200+ industry AI projects across 30+ sectors including finance, manufacturing, energy, and government, with cumulative project value exceeding RMB 1 billion. Deep expertise in extracting data from customer scenarios, building models, delivering results, and feeding failure cases back into the next training cycle.

Two Platforms Driving Global Leadership

On September 16, the WorldArena 2.0 global finals rankings were announced. Xirang Kaiwu achieved #1 globally in Trajectory Accuracy (Track 1: Video Quality) and Top 3 globally in Physical Adherence—the former measuring precise trajectory reconstruction given action policies, the latter directly assessing compliance with physical laws. Both are key indicators of physical modeling capability.

This performance is supported by two infrastructure platforms driving LPM's continuous evolution:

XR Data Platform: Designed for physical AI R&D, covering multimodal data ingestion, cleaning, annotation, quality inspection, version management, and release. It maintains cross-embodiment, cross-model neutrality, converting real-world robot trajectories, human demonstrations, and failure cases into trainable, traceable, and iterable high-quality data assets.

XR Evaluation Platform: Enables a complete workflow from scenario construction, quality inspection, policy execution, to report generation. It supports reusable evaluation scenarios and allows saving and reviewing execution processes and reports. The platform connects simulation scenario building, model inference, and result visualization into an automated closed loop. Failure cases can be located, supplemented, and re-verified, continuously feeding back into the next training cycle—creating a flywheel effect where evaluation drives improvement.

One Foundation for All Physical Scenarios

Xirang Kaiwu aims to become the infrastructure company for the physical intelligence era. Its commercialization path follows a clear layered logic:

• L0: General-purpose physical foundation model providing reusable physical capabilities
• L1: Domain-specific models fine-tuned on industry data
• L2: Application models distilled for specific scenarios, completing the delivery loop

The first step is to close the loop in robotics—partnering with hardware manufacturers and scenario owners to accumulate data, build evaluation tools, and create verifiable delivery templates. The second step is to replicate the same foundation to broader physical world applications: closed-loop simulation for autonomous driving, extreme weather prediction and long-term evolution simulation, drug molecule generation and property prediction. These challenges, previously reliant on specialized models and brute-force computation, are fundamentally about "understanding physical laws and predicting state evolution"—precisely LPM's domain.

The window for large-scale pre-training of physical AI is now open. Teams with the genuine capability to "train physical AI models from the ground up" are extremely rare. Xirang Kaiwu, anchored in first principles, following the path of native pre-training, and validated through industrial closed loops, is delivering the "physical foundation model" answer for this era.