A young Chinese physical-AI company called XIRRA says it has completed seed and angel financing worth several hundred million yuan, placing a reported US$500 million valuation on a business built around what it calls a native physical foundation model. The announcement comes as investors continue to fund companies seeking to give AI systems a more practical understanding of movement, contact, and changing real-world environments.
The September 23 dispatch by CLS identifies the company as 息壤开物, rendered as XIRRA in the reporting. It says Dunhong Asset led the financing, with several other investors participating, although the source record contains a discrepancy over one co-investor’s name. CLS calls the total proceeds 数亿元, meaning several hundred million yuan, and does not disclose an exact amount or how the funding was split between the seed and angel rounds.
That distinction is essential. The reported US$500 million figure is a valuation, not the amount raised. The sources do not publish the investment terms, ownership stakes, or closing dates. They do, however, describe a broad use of proceeds: pretraining a physical-AI model, building real physical-interaction data, hiring core research staff, and validating the technology in scenarios outside the lab.
XIRRA’s announcement sits within the wider physical-AI discussion. EastFrontier’s Kinetix coverage and its report on JD’s world model offer related historical context. XIRRA’s own technical claims must remain separate from those stories.
Two early rounds, one broad technical ambition
CLS says the company has completed a seed round and an angel round in succession. Dunhong Asset is named as lead investor. The primary report also lists Huakong Fund, Sanhua Holdings, Ginkgo Valley Capital, Zeran Capital, Benjian Fund, Tian Shi Ji Shi, and Biaopu Investment as follow-on investors.
Other reports agree on the lead investor but render one of the follow-on names differently, listing Shanghai Angel Club rather than Tian Shi Ji Shi. Since the sources do not resolve that difference, it would be misleading to present a definitive combined investor list. The safer approach is to name Dunhong Asset as lead investor and either omit the disputed name or identify each list by source.
The funding itself signals that physical AI is being treated as a model-development problem as well as a hardware problem. Companies in this area often need to collect data from interactions that are expensive and slow to reproduce: grasping objects, moving through space, reacting to force, and recovering from unexpected changes. A model trained only on internet text and images may have little direct experience of those interactions.
XIRRA says it will use the financing to build exactly that missing layer. It plans large-scale pretraining for native physical-AI models and more real physical-interaction data. The company also says it will use the money for research hiring and scenario validation. These are planned uses of proceeds, not completed achievements, but they show where the company sees its bottleneck.
LPM and the problem of evaluating physical understanding
The company’s model is called LPM. A report carried by 10jqka and Investment界 expands the label as Large Physics Model and describes LPM as supporting physical-law understanding, state prediction, long-horizon planning, and closed-loop execution. The primary CLS dispatch confirms only the LPM name, so that expansion and technical description require the secondary-source attribution.
CLS and TMTPost say LPM recently placed first globally for trajectory accuracy and in the global top three for physical compliance at the WorldArena 2.0 global finals. The reports do not provide a public leaderboard, score values, methodology, competitor list, or independent verification. Those results should remain company or media-reported benchmark claims, not evidence that LPM has solved real-world robotics.
A 36Kr analysis adds an important qualification. It says the result concerned a Track 1 video-quality subtask and should not be treated as a comprehensive demonstration of robot performance. That distinction is easy to overlook. A system can perform well in a simulation or a narrowly defined benchmark while still struggling with the unpredictable surfaces, forces, delays, and safety constraints of a physical deployment.
The gap between a benchmark and a working robot is why the company’s data plans matter. Physical foundation models will need carefully recorded interactions, diverse environments, and ways to check whether a prediction corresponds to what actually happens. The more a model is asked to plan long sequences, the more damaging small errors in physical assumptions can become.
The company and the physical-AI field
XIRRA was founded by Li Yin and Zhang Hanwang, according to CLS and TMTPost. CLS describes Li as a former chief technology officer for Huawei Cloud’s large-model business and Zhang as a professor at Nanyang Technological University. The source material does not establish additional executive titles, legal roles, or an official English company profile, so the names and backgrounds should not be expanded beyond that reporting.
The company is entering a field with no settled vocabulary. Some participants speak of world models, others of embodied intelligence, and XIRRA uses the language of physical foundation models. The common challenge is to build software that can predict consequences in the physical world well enough to support useful action.
That will be a difficult commercial test. The navigated sources do not state regulatory approval, securities registration, or other legal status, and they do not disclose financing terms, ownership percentages, or exact closing dates. The financing is an early event, not proof of market adoption.
Still, the combination of several hundred million yuan in early financing and a reported US$500 million valuation shows investors’ willingness to support new approaches to physical AI. The rationale is straightforward: if robots and autonomous systems are to work in more varied environments, they will need better ways to represent forces, trajectories, and the consequences of actions.
XIRRA’s next work will be more important than its first valuation. The company says it will spend the capital on model pretraining, real interaction data, research talent, and scenario validation. Whether those ingredients produce a broadly useful physical model remains unproven. For now, the funding provides an early marker in China’s increasingly competitive effort to make AI reason beyond the screen.