Huawei Large-Model Veterans Launch Xirang Kaiwu to Build Physical-World Foundation Model

The rounds were led by Dunhong Asset Management, with Huakong Fund, Sanhua Holding, Ginkgo Valley Capital, Zeran Capital, Benjian Fund, Angel Cornerstone, Biaopu Investment and other institutions participating, QbitAI reported. The company is pursuing an answer to a question it says remains unresolved in embodied intelligence: after large language models, does the physical world still need a foundation model?

Xirang defines LPM as a model pre-trained from massive video, robot trajectories and real interactions before being adapted to different embodiments and tasks. It contrasts this with LLMs, which learn knowledge humans have already abstracted into language, code and formulas. In Physical AI, which the company calls AI 2.0, the world is continuous, changeable by action and filled with alternative possibilities, while physical reality provides only pixels, sounds, positions and force changes rather than ready-made abstract representations. The model must discover useful representations from billions of observations, actions and feedback.

According to the company, the limitations of the VLA route are a major reason LPM is necessary. VLA has shown that visual and language information can be mapped into robot actions, but deployments still require additional training for specific bodies and tasks. If robots are to enter homes and public-service settings, models need general capabilities and must retain learned physical knowledge when entering new environments. Xirang positions LPM as that capability base.

The company points to two early signals. In World Arena 2.0 Track 1, its RiXin model ranked first in Trajectory Accuracy and second in Physical Adherence, according to QbitAI. Trajectory Accuracy measures whether predicted object trajectories are accurate; Physical Adherence checks whether generated future states obey basic physical laws. Xirang also said it has seen initial performance improvements as data and training scale increase, consistent with its hypothesis that physical intelligence may have a scaling path.

Li Yin, founder and CEO, graduated from Tsinghua University and was formerly CTO of Huawei Cloud's large model business and director of its industry large-model development department. QbitAI reported that she led a 400-person team that trained an 8B video model in two months on a 10,000-GPU cluster using a million-hour video dataset. On the industry side, she was responsible for launching embodied-intelligence innovation centers in several Chinese locations and delivered more than 200 customers across 30 industries, including autonomous driving, mining, finance, healthcare and weather, generating hundreds of millions of yuan in annual revenue.

Zhang Hanwang, co-founder and chief scientist, was formerly Huawei's chief multimodal scientist and is a professor at Nanyang Technological University's College of Computing and Data Science and an AI President's Chair Professor. He has published 228 papers, been cited more than 42,000 times and has an H-index of 81, with work focused on causal machine learning, multimodal understanding and model generalization. The report described him as the top Chinese scholar in causality. He was named among IEEE's top 10 AI researchers to watch and received Singapore's President's Science and Technology Award for Young Scientists.

Xirang's technical approach begins with long-horizon tasks in physical settings. A robot cleaning a table must recognize objects, decide gripping order, plan paths and adjust strategy after every action. If there are b choices at each moment over H steps, possible paths grow as b^H, producing combinatorial explosion. The company's autoregressive architecture is tied to that problem. Zhang has described the universe as a continuously unfolding autoregressive process in which each state follows prior evolution and combines with current actions to produce the next state. This draws on Markov processes and Bellman recursion: if the current state contains enough information, long histories can be compressed into the present, and future value can be decomposed into immediate reward plus value from the next state. The model does not need to store every past frame or enumerate every future, but can act from the current state and continue as the state changes. Xirang compares this to chess: a player judges the current position, chooses a move and continues rather than enumerating all possible games.

The LPM uses a Unified AR Transformer as its backbone, allowing an autoregressive model to break complex processes into continuous conditional predictions. Diffusion is an optional module for rendering future states as continuous video, which the report compares to script and storyboard work versus cinematography and post-production. The company divides the LPM into six foundations: unified tokens, unified model architecture, reinforcement-learning algorithms, data systems, evaluation systems and engineering infrastructure. Unified tokens convert visual, action, tactile and force information into shared representations; the unified architecture organizes state understanding, reasoning, action and future prediction in one model; reinforcement learning updates policies from action outcomes; the data system supplies learning material; evaluation checks whether the model has learned physical laws or only surface correlations; and the engineering base determines whether the method can run at sufficient data and compute scale. These form a loop: data is converted into unified representations, the model generates actions and predicts consequences, actions are tested in simulation or reality, rewards and evaluation turn results into feedback, and success and failure return to training.

First-phase data will come from internet physical videos, ego-centric first-person data, UMI demonstrations, simulation and synthetic data, and public robot trajectories and real-machine feedback. Xirang says its past commercialization experience gives it opportunities to enter real industrial scenes for data and to run a data flywheel. Its product plan has three levels: L0, a general physical foundation model for cross-task, cross-scene and cross-embodiment abilities; L1, domain models that learn objects, processes and constraints in sectors such as industrial manufacturing; and L2, applications that turn domain capabilities into deployable robot tasks. The company plans to unveil XIRRA v0.1 later this year.