Huawei AI Stars Co-found Startup Xirang, Secures Hundreds of Millions in Funding

Xirang Kaifu announced the successive completion of hundreds of millions in seed and angel funding, led by Dunhong Asset, with participation from Huakong Fund, Sanhua Holding, Ginkgo Valley Capital, Zeran Capital, and other institutions, reaching a valuation of $500 million. This capital is directed toward an experiment on a physical foundation model that has not yet completed pre-training.

Behind the company are two people known as Huawei's large-model twin stars: founder and CEO Li Yin, former CTO of Huawei Cloud Pangu Large Model and Director of the Industry Large Model Development Department; co-founder and Chief Scientist Zhang Hanwang, former Huawei Chief Multimodal Scientist and professor at Nanyang Technological University. Their capabilities correspond respectively to the engineering-industry side and the algorithmic research side of Huawei's large-model system.

The fact that capital is willing to bet before the model has completed pre-training itself shows this is a non-consensus answer. The question Xirang aims to answer is: after large language models, does the next stage facing the physical world still require a foundation model?

In Xirang's definition, LLM represents AI 1.0, mainly learning knowledge that has already been abstracted by humans into language, code, and formulas; Physical AI, which it calls AI 2.0, faces a world that changes continuously, can be altered by actions, and behind each action lies countless possibilities. The physical world does not provide ready-made abstract representations; it only provides pixels, sound, position, and changes in force.

Xirang believes the VLA route has limitations: although visual and language information can be mapped to robot actions, in actual deployment many models still require additional training for specific embodiments and tasks. If robots are to enter homes and public service scenarios, models need general capabilities and must retain learned physical knowledge in new environments. LPM hopes to provide exactly such a capability foundation.

To this end, Xirang chose an autoregressive architecture, with a Unified AR Transformer as the backbone, decomposing complex processes into continuous conditional predictions; Diffusion serves as an optional module to complete visual generation when future states need to be rendered into continuous video. Its theoretical intuition can be traced back to Markov processes and Bellman recursion: the past need not be saved frame by frame, and the future need not be exhaustively enumerated at once; the model only needs to choose an action in the current state, then continue inferring based on the new state.

Xirang disclosed two optimistic signals. First, its model ranks first in Trajectory Accuracy and second in Physical Adherence in World Arena 2.0 Track 1; the former examines whether the model's predicted object motion trajectories are accurate, while the latter focuses on whether the generated future states conform to basic physical laws, both directly related to the problem LPM aims to solve. Second, according to Xirang, as data scale and training scale increase, model performance has shown an initial upward trend, and the team believes it is worth continuing.

But extending this technical curve is not easy. Once a ten-thousand-GPU cluster is running, it costs hundreds of millions per month, and few can sustain progress under such consumption. Funding reserves are one aspect; judgment about team organization and technical route is another.

Li Yin's experience spans foundation model training and industrial delivery: she once led a 400-person team, relying on a million-hour video dataset, to train an 8B video large model in 2 months on a ten-thousand-GPU compute cluster; on the industry side, she was responsible for landing embodied intelligence innovation centers in multiple domestic locations, delivering to over 200 customers across 30 industries including intelligent driving, mining, finance, healthcare, and meteorology, generating hundreds of millions in annual revenue. These experiences form the background for her judgment on the Physical AI route.

The difficulty of a physical foundation model is no less than that of an LLM. Take clearing a table as an example: the robot must first recognize objects, then determine the grasping order, plan movement paths, and continuously adjust strategy based on environmental changes after each action. If there are b choices at each moment and completing the task requires H consecutive decision steps, the possible paths grow to b^H; if there are 10 choices per step and 20 consecutive decision steps, the theoretical number of paths explodes to an unimaginable order of magnitude. This is the combinatorial explosion in long-horizon decision-making.

Zhang Hanwang's research is directly related to this problem. In the physical world, statistical correlation alone is insufficient to support stable action: an object's movement may come from robot action, or from gravity, collision, or external disturbance; only by understanding the causal relationship between actions and outcomes can the model make more reliable judgments; the same applies to generalization—if it only memorizes object appearances and robot poses in training data, capabilities may fail when materials, lighting, or embodiments change. His research on causal learning and debiasing can precisely reduce the model's reliance on surface correlations.

Based on this framework, Xirang splits LPM into six mutually cooperating underlying capability foundations: Unified Token is responsible for converting heterogeneous information such as vision, action, touch, and force into representations that can be processed together; Unified Architecture is responsible for organizing state understanding, reasoning, action, and future prediction within the same model; reinforcement learning allows the model to update strategies from action outcomes; the data system provides learning materials; the evaluation system tests whether the model has learned true physical laws or false surface relationships; the engineering foundation determines whether this approach can run at sufficiently large data and compute scales. This forms a closed loop: data is converted into unified representations, the model generates actions and predicts consequences, actions enter simulation or the real environment for testing, rewards and evaluations turn results into feedback, and the experience of success and failure re-enters training.

According to Xirang's plan, the first phase of data will come from internet physical videos, Ego first-person data, UMI demonstrations, simulation and synthetic data, as well as public robot trajectories and real-machine feedback. Its rich past commercialization experience gives it the opportunity to enter real industrial scenarios to obtain data and quickly spin up a data flywheel.

In product planning, Xirang divides model capabilities into three levels: L0, L1, and L2: L0 is a general physical foundation model, aiming to form underlying capabilities across tasks, scenarios, and embodiments; L1 is a domain model, which after entering specific domains such as industrial manufacturing still needs to learn the objects, processes, and constraints in the industry; L2 enters specific applications, converting domain capabilities into deployable robot tasks. Its training result XIRRA v0.1 is planned to debut later this year.

Whether a foundation model conforming to a Scaling Law truly exists in the physical world remains an open question. Xirang's ranking results and initial upward trend come from its own disclosure and await third-party verification; the model has not yet completed pre-training, and XIRRA v0.1 has not yet debuted. Whether the criteria for judging the value of physical models will be updated can only be awaited.