The lecture hall at Tsinghua was standing-room only, and the speaker was not a visiting celebrity but one of the school's own professors. Tang Jie (唐杰) has spent his career moving between the classroom and the company he helped build. On that day he told students the chatbot era was basically over — and that the interesting work was only beginning. For a man whose models now trade on the Hong Kong exchange, the message was a strategy memo disguised as a seminar.
The professor who stayed in the lab
Tang is a professor in Tsinghua University's department of computer science and, alongside CEO Zhang Peng (张鹏) and chairman Liu Debing (刘德兵), one of the three figures who steer Zhipu AI (智谱). He is also the company's chief scientist and a fellow of both IEEE and AAAI. Where many startup founders flee academia for the runway, Tang kept one foot in the seminar room — a posture that shapes how Zhipu talks about research.
His roots are in Tsinghua's Knowledge Engineering Group (KEG), one of China's oldest labs for natural-language processing and knowledge graphs, founded in 1996. Tang had already built AMiner, an academic search engine, before the model era arrived.
From AMiner to GLM
Zhipu was founded in 2019 as a spin-out from the KEG lab. The through-line was the GLM (通用语言模型) architecture — Tang is credited as the lead on GLM-130B, a 130-billion-parameter bilingual model released in 2022. In March 2023 the company open-sourced ChatGLM-6B, one of the earliest openly available Chinese chat models, and a community of developers formed around it.
Under Tang's technical direction the GLM family expanded into multimodal, coding, and agent models. The research culture — publish, open-source, iterate — distinguished Zhipu from peers that treated models as closed products.
The Hong Kong listing
In January 2026 Zhipu became the first publicly listed large-model company anywhere, trading on the Hong Kong Stock Exchange under code 2513.HK at an IPO price of HK$116.20. The offering raised about HK$4.35 billion (≈ US$560M) and valued the company at roughly HK$51.1 billion (≈ US$6.6B, RMB 51.8B) at listing, up from a pre-IPO valuation of about RMB 24.38 billion (≈ US$3.4B).
The financials show a company growing fast but spending faster. Revenue climbed from 57.4 million yuan in 2022 to 312.4 million yuan in 2024 — a compound annual growth rate near 130% — yet net losses widened to 2.96 billion yuan in 2024 as compute and research bills mounted. Zhipu's bet is that being first to public markets buys it the capital to keep training.
Where Tang thinks models go next
In a Tsinghua lecture reported in Chinese media in late 2025, Tang argued that the "chat" phase of AI is largely settled and that 2026 should be about models that can remember, improve themselves, and carry out long-horizon tasks on their own. He framed this as the real path to AGI (通用人工智能): not cleverer conversation, but autonomous work.
In published comments he has also said the first principle of model applications should not be inventing new apps, but replacing human roles — that 2026 would be the year AI begins taking over distinct jobs. These are Tang's own framings, reported in Chinese-language outlets; we have paraphrased rather than quoted verbatim because the originals are not in English.
Why a chief scientist's roadmap matters
Tang's move from the chat phase to "models that work" is a research leader's version of the same founder dilemma: capability demos are cheap, reliable autonomous work is hard. Zhipu's open-source habit lets outside developers verify the claim instead of taking it on faith — the GLM lineage is inspectable on Hugging Face, not locked in a demo video. That transparency is part of why the company reached public markets first. For readers, the takeaway is to watch whether Zhipu's agent and long-horizon models actually ship as tools people pay to use, rather than as benchmark wins. A chief scientist's roadmap is only as good as the products that follow it.
Honest limitations
Corporate facts (founding year, Hong Kong listing date and code, IPO price, proceeds, valuation, revenue and loss figures) are sourced from Caixin, Caijing, and Guancha reporting on Zhipu's prospectus and listing. Tang's biographical details come from company disclosures and Chinese tech-media profiles. His forward-looking remarks about model direction are paraphrased from Chinese-language lecture coverage (ZAKER and similar); we did not find an authoritative English transcript, so the wording is our rendering, not a direct translation of a verified English quote. We have not independently audited Zhipu's financial statements beyond what the prospectus summaries disclose.
What readers can do now
- Read Zhipu's Hong Kong prospectus (HKEX 2513) to see how a research-heavy Chinese lab presents its economics to public investors.
- Try the open GLM models on Hugging Face or Zhipu's platform to judge the technical lineage Tang's team built from GLM-130B to today.
- Follow Tang Jie's published talks for a research-led view of where Chinese model development is heading after the chatbot phase.