--- title: "Zhipu's Tsinghua founders chose a homegrown model architecture — and took it to Hong Kong" date: 2026-09-28 category: Builders site: NeuroAI canonical: https://neuroai.site/a/na-people-zhipu-founder language: en --- # Zhipu's Tsinghua founders chose a homegrown model architecture — and took it to Hong Kong > Spun out of a Tsinghua University lab in 2019, Zhipu built the GLM large model (大模型) family and became the first AGI-base-model company to list publicly. The founders — CEO Zhang Peng and chief scientist Tang Jie — are the academic-and-engineering pair behind China's "OpenAI" comparison. Most Chinese AI labs in 2019 were either copying or fine-tuning someone else's model. A small team walking out of a Tsinghua University laboratory decided to write their own architecture instead — and bet the company that it would still matter seven years later. ## A lab, not a startup, first Zhipu (智谱) was founded in 2019 in Beijing as a technology-transfer spinout from Tsinghua University's Knowledge Engineering Group (KEG) lab. The founding team reads like a faculty roster: CEO Zhang Peng (张鹏), chief scientist Tang Jie (唐杰) — a Tsinghua professor who led the GLM-130B project — and chairman Liu Debing (刘德兵). Tang Jie is the academic face; Zhang Peng runs engineering and commercialization. The defining early choice was architectural. Instead of the dominant GPT-style autoregressive stack, Zhipu backed GLM, a general language model built on an autoregressive blank-infilling pretraining paradigm. In 2021 the company released the GLM framework; in 2022 it open-sourced GLM-130B, a hundred-billion-plus-parameter bilingual model and one of the first open Chinese foundational models at that scale. ## The "Chinese OpenAI" label Zhipu leaned into research-mode product discipline rather than consumer-app marketing. In 2023 it released ChatGLM-6B, an early open-source dialogue model that drew a large GitHub and Hugging Face following. It followed with the GLM-4 family and, by late 2025, GLM-4.7, which the company says reached the top of several open-model benchmarks. The architecture has also been adapted to more than 40 domestic chips — a deliberate sovereignty move in a market wary of export controls. That breadth, Zhipu argues, is what keeps the model runnable if overseas hardware or APIs become unavailable. ## The listing that made it a number On 8 January 2026 Zhipu listed on the Hong Kong Stock Exchange under code 02513.HK, pricing at HK$116.20 and raising more than HK$4.3 billion (≈ US$550M). It billed itself as the world's first publicly listed company built on a general AI (AGI) base model — "the global large-model first stock." Market cap opened around HK$52–57 billion (≈ US$6.7–7.3B). The financials behind the debut show a research-first company still buying growth with losses. Revenue was 312.4 million yuan (≈ US$44M) in 2024 and 191 million yuan (≈ US$27M) in the first half of 2025, up 325% year on year. R&D spending was 2.195 billion yuan (≈ US$310M) in 2024 alone, and cumulative R&D across 2022–2025 reached about 4.4 billion yuan (≈ US$620M); roughly 74% of staff are researchers. ## Why the founders matter more than the IPO The interesting question is not the market cap but the bet. Zhang Peng and Tang Jie kept Zhipu on a self-developed path when cloning a foreign model would have been faster and cheaper. That bought independence — GLM runs on domestic silicon — but also a heavy cost burden and a long road to profitability. Tang Jie, as chief scientist, anchors the academic credibility; Zhang Peng, as CEO, has to turn a research lab's output into a business that pays for the next model. The pairing is Zhipu's core asset and its core risk: if the architecture stops leading, the whole premium compresses. ## What readers can do now - **If you build on Chinese models**, test GLM on a domestic-chip endpoint and compare latency against a Western model; the 40+ chip adaptation is the part worth verifying yourself. - **If you report on AI**, read Zhipu's R&D-to-revenue ratio as a proxy for how much China's model builders are spending to stay independent. - **If you invest**, separate the "first stock" narrative from the loss trajectory; the listing funds R&D, it does not end the burn. ## Honest limitations Facts on founding (2019, Beijing, Tsinghua KEG spinout), leadership (CEO Zhang Peng, chief scientist Tang Jie, chairman Liu Debing), architecture (GLM, 2021 framework, GLM-130B open-sourced 2022, ChatGLM-6B 2023, GLM-4.7 late 2025) and 40+ domestic-chip adaptation come from Zhipu materials and China Daily / 证券时报 (Securities Times) coverage. HKEX listing (8 Jan 2026, code 02513.HK, HK$116.20 issue price, >HK$4.3B raised, ~HK$52–57B market cap, "global large-model first stock") is from China Daily and 证券时报. Revenue (312.4M yuan in 2024; 191M yuan 1H2025, +325%) and R&D (2.195B yuan in 2024; ~4.4B yuan cumulative 2022–2025; 74% R&D staff) are from the IPO prospectus as reported by 证券时报. Tang Jie's role as GLM-130B principal investigator and Tsinghua professor is widely reported. "Chinese OpenAI" is a media description, not a company claim. Listing valuation is a market figure; RMB converted at ~7.1/US$ and HKD at ~0.128/US$. Not investment advice. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-people-zhipu-founder Free to quote with attribution and a link to the original.