Ex-Huawei Pangu CTO raises hundreds of millions for a physics model
Two of Huawei's best-known large-model people have a new company, and it isn't Huawei's. Shenzhen's XIRRA has raised several hundred million yuan to build a foundation model for the physical world — and the framing around it needs one correction.
Li Yin, the former CTO of Huawei Cloud's Pangu large-model team, left to found XIRRA (Shenzhen Xirrang Kaiwu Technology), and the company has now closed seed and angel rounds worth several hundred million yuan in total. Dunhong Assets led, with Huakong Fund, Sanhua Holding, Ginkgo Valley Capital and others following, according to Jiemian News, 36Kr and TMTPost, which reported the raise the same day. The corporate record shows the entity registered in Shenzhen's Bao'an district on 1 July 2026 with ¥1 million of registered capital — the money sits in the round, not the shell. Li's co-founder and chief scientist is Zhang Hanwang, an NTU professor with more than 228 papers and a citation count above 42,000, who is also Huawei's current chief scientist for multimodal AI.
The product is the reason to care. XIRRA is building an LPM — a Large Physics Model — pitched as a base model for embodied AI rather than for text. The architecture stacks an L0 base, L1 domain and L2 scenario layer, pairs a world model with a policy in one objective, and borrows the Bellman equation from reinforcement learning to tie them together. That is the same thesis Nvidia's Cosmos and a wave of Chinese world models are chasing: a model that understands how objects move and interact, because a robot that only understands sentences cannot pick anything up. Li's own account of the pace is the interesting detail — a video model trained on 1 million hours of footage on a 10,000-card cluster in two months, per a forum appearance days before the funding news.
Two corrections belong in the same breath as the announcement. First, Huawei is not involved — it neither invested nor is a party to the company; "Huawei's twin stars" is media framing of two people with Huawei on their résumés. Zhang is described in the company's disclosure as will become co-founder, which is an odd tense for someone still in a Huawei title. Second, the benchmark claim does not reproduce. XIRRA says its world model ranks first globally on Trajectory Accuracy in WorldArena 2.0, the Tsinghua-led evaluation; the public leaderboard entry for XIRRA-WM shows rank 3 of 81 on that metric and rank 5 on Physical Adherence. That could be a snapshot difference between the September 16 final board and today's — but as published, it is company-claimed, not verified. We are printing it as the former.
Alibaba Security's 27B model now tops CyberGym's model-focused track, beating systems several hundred times its size. XekRung-1.5-27B-Preview — a fine-tune of Qwen3.8-27B — posts an 88.9 percent success rate on the Berkeley RDI-run security benchmark, ahead of GPT-5.5-Cyber at 85.6 percent, Gemini 3.8 Flash Cyber at 86.26 percent and Claude Mythos Preview at 83.1 percent. The uplift is the story: the Qwen base model manages 54.51 percent on the same task, so the fine-tune adds roughly 34 points.
Read it with two caveats. The leaderboard's rows are self-submitted by each team and the board itself notes agent runs are stochastic, so scores vary; and "top" means the model-focused track only — on the agent-focused track Alibaba's own agent entry sits mid-pack at 85.34 percent while another team leads at 98.47 percent. The "beats trillion-parameter models" line in the Chinese coverage is also framing: nobody has disclosed the parameter counts of the frontier systems it is being compared against.
What to watch: whether XIRRA publishes enough of LPM's evaluation to reproduce the WorldArena claim, and whether any lab outside China picks up the 27B security fine-tune as a cheap alternative to a frontier model for red-teaming.
Would you trust a 27-billion-parameter model to red-team your own systems, or does security need the biggest model you can afford? Tell us in the comments.
