Tang Jie (唐杰) started with a notebook, a desktop and a PhD bonus of RMB 20,000 (US$2,800 / HK$21,800). Twenty years later the knowledge-engineering lab he ran at Tsinghua University (清华大学) had produced the GLM architecture behind Zhipu AI (智谱), a company that listed in Hong Kong on 8 January 2026 at a market value near HK$53 billion (US$6.8B).
Key takeaways
- Tang, a Tsinghua professor and ACM/AAAI/IEEE Fellow, founded the AMiner (学者大数据) academic platform in 2006 and co-founded Zhipu AI in 2019 with Zhang Peng (张鹏) and Liu Debing (刘德兵); he is chief scientist.
- He designed GLM (General Language Model, 通用语言模型), an "autoregressive blank infilling" architecture; GLM-130B (2022) was the only Asian model in Stanford's 30-model evaluation that year.
- As academic vice-dean of the Beijing Academy of Artificial Intelligence (BAAI, 北京智源人工智能研究院), he led the WuDao (悟道) 1.0/2.0 models in 2021, exceeding 1.75 trillion parameters.
- In June 2025 Tang stepped down from Zhipu's board, returned to Tsinghua and became the company's largest natural-person shareholder; Zhipu listed on HKEX on 8 Jan 2026 as the "world's first AGI base-model public company."
- Zhipu's prospectus shows R&D spend of about RMB 4.4 billion (US$615M / HK$4.8 billion) from 2022 through H1 2025; Frost & Sullivan ranked it first among China's independent general large-model developers by 2024 revenue.
From a RMB 20,000 bet
Tang did not begin with a company. In 2006, as a young Tsinghua faculty member, he launched AMiner, a scholarly knowledge graph that mapped global researchers, papers and collaborations. The startup legend is modest: three people, two machines, and a research grant drawn from his PhD bonus. Two decades on, AMiner spans more than 220 countries and regions and serves government bodies and research institutions as a science-intelligence backbone.
That project taught Tang a habit he carried into models: treat knowledge as a first-class object, not a side effect of statistics. AMiner was, in effect, a knowledge graph long before large language models made "retrieval" fashionable.
The GLM wager
In 2019 Tang and his Tsinghua collaborators turned the lab's accumulated methods into a company, Zhipu. The technical bet was unusual: instead of cloning the Transformer or following BERT and GPT separately, Tang's team built GLM around "autoregressive blank infilling," a single formulation that unified understanding and generation. That decision gave Zhipu a self-owned architecture — a rarity among Chinese model firms that mostly mirrored overseas designs.
The proof points followed. GLM-130B, released in 2022, performed on par with GPT-3's base and was the only Asian model in Stanford's 30-model evaluation that year. ChatGLM and its open ChatGLM-6B then climbed the Hugging Face trending list for weeks, and Zhipu reported the base model downloaded by more than 1,000 organizations across 70-plus countries. According to Frost & Sullivan, cited in Zhipu's prospectus, the firm led China's independent general large-model developers by 2024 revenue.
The BAAI years and WuDao
Tang also worked the national-research track. As academic vice-dean of BAAI and head of its WuDao technical committee, he oversaw the WuDao 1.0 and 2.0 models released in 2021, with WuDao 2.0 exceeding 1.75 trillion parameters — a global record at the time. Holding both a company chief-scientist role and a flagship state-lab role let him test ideas across academic and product settings, a dual position few in the field matched.
The knowledge-enhancement philosophy ran through both. Rather than simply scaling data, Tang argued for coupling neural models with symbolic knowledge — graphs and rules — so the system could be "grounded" rather than merely fluent. That line is now Zhipu's differentiating story against pure-scale competitors.
The listing and the long game
In June 2025 Tang left Zhipu's board, returned to Tsinghua, and became the company's largest natural-person shareholder — a move that kept the professor tied to the firm without running it. On 8 January 2026 Zhipu listed on HKEX (02513.HK) at HK$116.20 per share, raising more than HK$4.3 billion and opening at a market value around HK$53 billion (US$6.8B). The offering was oversubscribed more than 1,159 times in the public tranche. In an internal letter that day, Tang said the new GLM-5 would let "AI assist users in accomplishing more real-world tasks" — his phrasing for the applied turn the company is taking post-listing.
What it signals
Tang's career is the strongest counter-example to the idea that Chinese models are derivative. A self-owned architecture, a knowledge-enhanced doctrine, and a lab-to-IPO pipeline all trace back to one professor's decision to keep building rather than borrow. For the field, Zhipu's listing suggests the capital markets now price original architecture and vertical depth above raw parameter bragging — and that academic founders can outlast the hype cycle by staying close to the science.
Honest limitations
This profile uses Tang's Tsinghua homepage, Zhipu's prospectus as reported by China Daily and Investing.com, BAAI's site, and secondary accounts from Baidu Baike and Zaker. The RMB 20,000 starting grant, the 1.75-trillion-parameter WuDao claim, and the HK$53 billion opening valuation are sourced from those outlets restating company or exchange figures; I did not retrieve the primary AMiner server logs, the WuDao training records, or the HKEX filing directly. The "only Asian model in Stanford's 30-model evaluation" and Frost & Sullivan revenue-ranking claims are company- or prospectus-cited and I could not independently re-verify them. Tang's "real-world tasks" remark comes from a China Daily report of his internal letter, not a transcript I accessed. Where a fact had a single source, I omitted it rather than assert it.
What readers can do now
- When comparing Chinese model firms, weight self-owned architecture and knowledge-enhancement strategy, not just benchmark scores.
- Track Zhipu's post-IPO GLM-5 deployments in finance, energy and manufacturing as the real test of the "knowledge-enhanced" claim.
- Read the HKEX-listed prospectus as a primary document — it is one of the few audited data sources in this otherwise private-heavy sector.