House · Tsinghua University AI

From the resumption of teaching in 1978 to the era of large models in 2026, Tsinghua University single-handedly nurtured China's AI academic system, and single-handedly produced half of the protagonists of today's Chinese LLM startup wave.

I. From Restoration to Foundation (1978–1990)

The story begins in 1978. In that year of "Reform and Opening Up," Tsinghua's Department of Computer Science resumed admissions and research, and AI quietly took shape as a sub-direction of computer science. Domestic AI was then in a phase of "groping in the dark" — academic ties almost severed during the Cultural Revolution had only just been rebuilt, and a generation of researchers born in the 1940s and 1950s, including Zhang Bo (张钹) and Ma Songde (马颂德), returned from Britain or the United States or caught up by self-study, bringing back to Beijing the seeds of pattern recognition, theorem proving, and expert systems. The Tsinghua CS department then operated out of a few cramped offices in the east wing of the main building. The VAX minicomputer was the department's most expensive asset — and one had to queue with a ticket to use it.

In 1985, Tsinghua established the State Key Laboratory of Intelligent Technology and Systems (智能技术与系统国家重点实验室), led by the CS department — the first state key laboratory in mainland China to bear the word "intelligence" in its name. Its first director was Professor Shi Chunyi (石纯一), with core directions covering knowledge representation, theorem proving, pattern recognition, and intelligent robotics. That same year, the National "863 Program" (High-Tech Research and Development Program) was launched, and AI was listed as a key direction within information technology. Tsinghua became the national flagship in the AI track of the 863 Program.

The lab's true soul was Zhang Bo (张钹) — a 1958 graduate of Tsinghua's Department of Automatic Control who had stayed on as faculty, working first on automatic control and artificial neural networks, and turning in the 1980s to theorem proving and heuristic search. From 1980 to 1982 Zhang visited the University of Illinois at Urbana-Champaign (UIUC), forging firsthand ties with the AI community abroad. On returning, he quickly carried the idea that "AI is an independent discipline" back to Tsinghua. In 1995 he was elected to the Chinese Academy of Sciences, among the earliest CAS academicians in Chinese AI. The "Quotient Space Theory" he developed with students such as Ma Shaoping (马少平) remains one of the few internationally recognized original contributions from native Chinese AI theory.

In the same period, Tsinghua's Department of Automation also set up a Pattern Recognition and Intelligent Control Lab, where professors such as Li Yanda (李衍达) and Sun Zengqi (孙增圻) worked on expert systems and robot control. In the 1990s Tsinghua sent multiple cohorts of young scholars to study in the United States, many of whom would later become important forces in Chinese AI both in industry and academia — a vein of talent export and return that has run from the 1980s to the present day.

II. The Sudden Rise of Yao Class (2004– )

2004 was a turning point in Tsinghua AI's history.

That year, Andrew Yao (姚期智) resigned his tenured chair at Princeton and returned full-time to Tsinghua. He is, as of 2026, the only Turing Award laureate among ethnically Chinese computer scientists — winning the prize in 2000 for "fundamental contributions to the theory of computation," especially pseudorandom number generation, cryptography, and communication complexity. Born in Shanghai in 1946, Yao earned his physics undergraduate degree at National Taiwan University in 1967, a Harvard physics PhD in 1972, and a CS PhD from Illinois in 1975. He had taught at MIT, Berkeley, and Princeton, and was a towering figure in computational theory.

After returning, Yao did one thing of far-reaching consequence: in 2005 he founded the "Software Science Pilot Class," renamed in 2007 to the Tsinghua Xuetang Computer Science Pilot Class, popularly known as "Yao Class" (姚班). Each cohort of about 30 students is selected from Tsinghua's incoming freshmen; courses are taught entirely in English; theoretical training is intensive; and undergraduates are heavily involved in research. Yao personally taught the core theory courses, transplanting Princeton's "apprenticeship + early research" tradition wholesale to Tsinghua.

Over the past dozen-plus years, Yao Class graduates have come to dominate Chinese student slots at top theoretical computer science conferences (STOC, FOCS) and have flowed in great numbers into AI. Representative figures trained in the Yao system include Danqi Chen (陈丹琦) (Princeton professor, top NLP scholar in the BERT era), Tengyu Ma (马腾宇) (Stanford professor), Rong Ge (鬲融) (Duke professor), Tiancheng Lou (楼天城) (co-founder of Pony.ai), Tang Wenbin (唐文斌) (co-founder of Megvii), Yin Qi (印奇) (CEO of Megvii), and Yang Zhilin (杨植麟) (founder of Moonshot AI, Yao Class 2011 cohort, Xuetang Physics Class).

In 2011, Yao founded the Institute for Interdisciplinary Information Sciences (IIIS) at Tsinghua as the graduate-school extension of Yao Class — the institute became the central stronghold of Tsinghua's research in theoretical computer science and AI theory. In 2019, building on IIIS, Yao opened the "Zhi Class" (智班, AI Pilot Class), expanding the Yao Class model from theoretical CS into AI and tilting Tsinghua's "elite small class" pipeline further toward AI.

In numerous public talks Yao has stressed that his goal in coming home was not merely to train a handful of top students, but to "grow a tree of computer science of our own on Chinese soil." The result largely delivers on that ambition — roughly one third of Yao Class graduates enter academia (many on faculty at top US schools), one third return to China to start companies or join leading tech firms, and the remaining third work in finance and quantitative trading (hedge funds).

The Yao Class system has not been without controversy — critics argue it reinforces a "theory above all" and "export-oriented" bias, channeling some students into a fixed track of "Yao Class → top-school PhD → Silicon Valley big tech or quant fund," rather than rooting them in China's domestic engineering problems. Even so, the existence of Yao Class gave China for the first time an undergraduate elite-training system that could compete head-to-head with MIT EECS and CMU CS — that much is fact.

III. The Long March of Knowledge Engineering (KEG and AMiner)

If Yao Class is the standard-bearer of Tsinghua AI theory, the Knowledge Engineering Group (KEG) is the citadel of applied AI at Tsinghua.

KEG is led by Professors Li Juanzi (李涓子) and Tang Jie (唐杰) of the Tsinghua CS department. Tang Jie — undergraduate at Central South University, Tsinghua PhD in 2006, retained as faculty — has worked on "academic social networks" since his doctoral days. In 2006 he launched the AMiner academic search engine project with the goal of "building a profile for every researcher and a knowledge graph for every paper." AMiner has since accumulated profiles of more than 300 million scholars and over 400 million papers, becoming the largest academic knowledge graph in the Chinese-speaking world and an unavoidable tool for Chinese AI researchers. Tang has accordingly become a leading figure in China's knowledge-graph community, and won the ACM SIGKDD Test-of-Time Award in 2018.

Tang has accumulated twenty years of work in knowledge graphs, social network analysis, and graph neural networks (GNN). Around 2020, his team began to pivot toward LLMs — a turn that would later spawn the most consequential industrial spin-off in Tsinghua's AI history.

The other key figure in KEG is Li Juanzi (李涓子), who from the 2000s devoted herself to Chinese knowledge graphs and domain ontologies, and was the de facto director of KEG in its early years. Her students cover almost every technical backbone of today's Tsinghua-affiliated LLM companies. In the early 2020s, KEG collectively swung its research direction "hard" from knowledge graphs toward LLMs — a turn that was no accident: knowledge graphs are themselves an attempt to "structuralize" the world, while LLMs are another path that "parameterizes" the world. The two roads began to converge after 2022.

IV. ChatGLM and Zhipu AI

In 2020, KEG joined Tsinghua CS Professor Huang Minlie (黄民烈) (a dialogue-systems specialist) and others to launch the GLM (General Language Model) pretraining project. GLM's technical hallmark is an "Autoregressive Blank Infilling" training objective that tries to unify the strengths of encoders and decoders — formally published at ACL 2022, it is one of the few original contributions at the architectural level among Tsinghua's LLM efforts.

In 2022, the team incubated Zhipu AI out of Tsinghua, with Tang Jie as CTO and Zhang Peng (张鹏, Tsinghua PhD 2017) as CEO. In August 2022, the team released GLM-130B — a 130-billion-parameter Chinese-English bilingual pretrained model, then the largest open-source Chinese LLM in the world, later published at ICLR 2023.

On 14 March 2023 — the very day OpenAI released GPT-4 — ChatGLM-6B was open-sourced. This was the first Chinese-developed conversational LLM with fully open weights and a commercial-use license, capable of running on a single consumer-grade GPU. It crossed 10,000 GitHub stars within a week and 20,000 within a month. The same year saw ChatGLM2, ChatGLM3, and afterwards GLM-4, GLM-4-Plus, GLM-Z (a reasoning model), and GLM-4V (multimodal) — making Zhipu AI the most academically pedigreed of "China's Six AI Tigers" (中国大模型六小虎).

By the end of 2025, Zhipu AI was valued at roughly RMB 20 billion. Its investors include Alibaba, Tencent, Meituan, Xiaomi, the National Social Security Fund, Beijing's Zhongguancun Science City, and the Beijing AI Industry Fund — making it a de facto national-team Chinese AI company. In October 2024, Zhipu AI was placed on the U.S. Department of Commerce's Entity List — the first Chinese LLM company to be sanctioned by Washington, an inverse confirmation of its strategic standing.

In 2024, Zhipu AI, along with Baichuan, 01.AI, Moonshot AI, MiniMax, and StepFun, was dubbed by Chinese media "China's Six AI Tigers" (六小虎) — the founding teams of the first four trace their core members to Tsinghua. Widening the lens to include underlying frameworks (early teams of OneFlow and PaddlePaddle) and embodied AI (Galbot 银河通用, Unitree 宇树科技), Tsinghua's coverage of Chinese AI talent extends further still.

V. The Tsinghua School of Generative AI

Beyond GLM, several other independent and important AI groups have emerged at Tsinghua —

• Zhu Jun (朱军), professor in the Tsinghua CS department, a leading figure in Bayesian deep learning and diffusion models. In 2023 his team incubated Shengshu (生数科技), releasing the video generation model Vidu, benchmarked against OpenAI's Sora.
• Sun Maosong (孙茂松), Tsinghua CS professor and a senior NLP scholar, leads the OpenBMB open-source LLM community and the CPM (Chinese Pretrained Models) series.
• Liu Zhiyuan (刘知远), Sun Maosong's student and CPM's chief technical lead. In 2022 he and lab-mates incubated ModelBest (面壁智能), releasing the on-device MiniCPM family. In 2024, the demo of MiniCPM-2.4B running on a 2 GB-memory device caught the international community's eye.
• Huang Minlie, expert in dialogue systems and affective computing, who in 2021 incubated the emotion-dialogue startup Lingxin AI (聆心智能).
• Ya-Qin Zhang (张亚勤), former Microsoft Global VP and former Baidu President, who joined Tsinghua in 2020 as Dean of the Institute for AI Industry Research (AIR), focused on autonomous driving, smart healthcare, and smart cities.
• Bowen Zhou (周伯文), former IBM Research Fellow and former head of JD AI Research, returned to Tsinghua's Department of Electronic Engineering in 2022 as professor, while concurrently serving as Director of the Shanghai AI Laboratory, leading work on general LLMs and embodied AI.

These groups cooperate and compete in roughly equal measure, forming the unique "hundred flowers in bloom" pattern of Tsinghua AI in the 2020s — theory (Yao Class / IIIS), knowledge graphs (KEG / Zhipu), Bayesian methods (Zhu Jun / Shengshu), NLP (Sun Maosong / ModelBest), industry (Zhang Yaqin / AIR), and platform (Bowen Zhou / Shanghai AI Lab) each have their own strengths.

A common feature of the "Tsinghua School" in generative AI is its emphasis on architectural innovation at the foundation level rather than mere application-stacking. Whether it is GLM's blank-filling pretraining objective, Vidu's diffusion-Transformer framework (U-ViT, proposed by Zhu Jun's team), or MiniCPM's "small model, big efficiency" path — each reflects an academic habit of "doing the research thoroughly before thinking about productization." Within the dominant Chinese internet logic of "fast iteration, scene capture," it forms a relatively unique current.

Zhu Jun's team's work on diffusion-model theory (early Diffusion Transformers, Chinese improvements on ScoreSDE) has repeatedly been a best-paper candidate at NeurIPS and ICML; Sun Maosong's team's work on LLM alignment and efficient training (OpenBMB, BMInf) has been cited by the international open-source community. The Tsinghua School's research stance leans more toward "systems science" than its Silicon Valley counterparts — a methodological line that traces back to the generation of Zhang Bo.

VI. The Tsinghua Family: Alma Mater of China's LLM Startup Wave

Survey the founder résumés of China's first-tier LLM startups from 2023–2025 and a striking fact emerges: almost every one of them is Tsinghua-affiliated:

Company	Founder(s)	Tsinghua Background
Zhipu AI	Zhang Peng, Tang Jie	Tsinghua PhD / KEG
Moonshot AI (Kimi)	Yang Zhilin	Tsinghua undergrad (Xuetang Physics), CMU PhD
ModelBest (面壁智能)	Liu Zhiyuan, Li Dahai (李大海)	Tsinghua PhD / NLP Lab
Shengshu (生数科技)	Zhu Jun, Tang Jiayu (唐家渝)	Tsinghua professor / PhD
Baichuan AI	Wang Xiaochuan (王小川)	Tsinghua undergrad (CS)
MiniMax	Yan Junjie (闫俊杰)	Tsinghua PhD (Automation)
Light Years Beyond (光年之外)	Wang Huiwen (王慧文)	Tsinghua undergrad (Electronics)

This phenomenon is exceptionally rare in China's startup history — a single university effectively monopolizing the founder pool of a trillion-RMB emerging industry. The reasons can be parsed into three layers of long-accumulated capacity at Tsinghua AI: first, theoretical training — Yao Class / IIIS supplied a top-tier mathematics-and-algorithms talent pool; second, engineering culture — the CS department's long-standing engineering ethos has produced compound founders who can code, ship, and raise capital; third, social network — Tsinghua's alumni network's integrative power across Chinese capital markets and industry resources far exceeds that of other universities.

A separate note on Yang Zhilin — trained in the Yao Class system at Tsinghua, then a CMU PhD under the prominent NLP scholar Ruslan Salakhutdinov, with classic Transformer-XL and XLNet papers (co-authored with Quoc Le) from the BERT era. Returning to China in 2023, he founded Moonshot AI; in 2024 Kimi Chat went viral on the strength of its "ultra-long context," with monthly actives breaking ten million in short order and valuation jumping to USD 3.3 billion — at one point the most imaginative consumer-facing Chinese LLM product.

Worth noting too is the late Wang Huiwen (王慧文) — a Tsinghua Electronics graduate and Meituan co-founder, who in early 2023 announced he would put USD 200 million of his own money into founding "Light Years Beyond." Just three months later, Meituan acquired the company; Wang himself stepped back from operations on health grounds — but the brief life of Light Years Beyond served as an icebreaker for China's LLM startup wave, called by many "the first shot fired in the LLM startup era."

Around these core teams sits a constantly replenishing roster of mid-career and rising faculty filling out Tsinghua AI's map — including Long Mingsheng (龙明盛, time series and causal discovery), Liu Yang (刘洋, machine translation), Ding Xiaohan (丁霄汉, vision and neural-architecture search), Ding Ning (丁宁, a CPM core author), Dong Yuxiao (东昱晓, KEG LLM direction), Li Zhongnan (李仲楠, multimodal), and others. Each has signature work in their own direction, together forming a dense "Tsinghua AI academic web."

VII. BAAI, Shanghai AI Lab, and the National Map

Tsinghua AI's reach does not stop at campus.

The Beijing Academy of Artificial Intelligence (BAAI, 北京智源人工智能研究院), founded in November 2018, is jointly built by the Beijing municipal government with Tsinghua, Peking University, and the CAS Institute of Automation. Tang Jie, Sun Maosong, and other Tsinghua professors are deeply involved. The "Wu Dao 2.0" model BAAI led released in 2021 reached 1.75 trillion parameters — at the time the largest multimodal model in the world.

The Shanghai AI Laboratory was established in 2020. From 2022, the Tsinghua professor (joint appointment) Bowen Zhou has served as its director, releasing such open-source projects as InternLM (书生·浦语), InternImage, and the OpenCompass evaluation suite.

Tsinghua professors shuttle between on-campus labs, BAAI, the Shanghai AI Lab, and their own companies as joint appointees or directors, weaving a dense network linking Chinese AI academia and industry.

The teaching infrastructure has continued to scale. In 2018 Tsinghua established the Institute for Artificial Intelligence, with Zhang Bo as honorary dean, Andrew Yao as chair of the academic committee, and Zhu Jun as deputy dean — China's first university-level AI institute. The Zhi Class opened in 2019 runs in parallel with Yao Class, forming a "twin elite class" structure for AI undergraduates at Tsinghua. At the graduate level, the CS department, Electronics, Automation, IIIS, and AIR all admit students; each year Tsinghua trains more than 500 master's and PhD students in AI directions — the largest pool of top AI talent in China.

By early 2026, Tsinghua AI sits firmly in the global top three on international academic indicators (CSRankings AI category), with paper output and top-conference acceptances second only to CMU and MIT. In industrial incubation it is, plainly, the Whampoa Military Academy of Chinese AI. A faintly ironic phenomenon — when Silicon Valley's top AI labs screen résumés today, the keywords "Tsinghua" and "清华" sit on roughly the same shelf as "Stanford" and "MIT." That shift took place between 2018 and 2023, almost in lockstep with the rise of China's LLM industry.

Worth a final note is the relationship between Tsinghua AI and Chinese AI policymaking. Zhang Bo, Andrew Yao, and Zhang Yaqin all serve long-term roles on the National Science and Technology Strategy Advisory Committee and the Ministry of Education's Academic Degrees Committee; Tang Jie, Zhu Jun, and Sun Maosong have repeatedly participated in the Ministry of Industry and Information Technology's and the Cyberspace Administration of China's LLM policy discussions. When China's Interim Measures for the Administration of Generative AI Services (August 2023) and the LLM filing system were issued, Tsinghua scholars' views entered the decision process directly — a tightness of coupling between academia and policy rare among the world's major AI nations, and a feature that distinguishes Tsinghua from Stanford, CMU, and MIT.

Historian's Note

Tsinghua is to Chinese AI what the Jixia Academy was to the Warring States — its excellence lies not in any one master, but in the coexistence of a hundred schools. Zhang Bo planted the root, Andrew Yao bestowed legitimacy, Tang Jie, Zhu Jun, and Sun Maosong each opened a vein, and Zhang Yaqin and Bowen Zhou stretched the reach into industry and government. This "scattered, polyphonic" ecosystem is unlike Silicon Valley's "giants + startup ecosystem" American paradigm, and unlike France's Mistral-style "a few stars hold up one banner" European paradigm — it is a unique academia-industry coupling cultivated over forty years by a single national university. Its strengths: a deep talent pool, plural schools of thought, and tight alignment with national strategy. Its hidden risks: the boundaries between research institute and company, university and government, academic publication and trade secret often blur — incubating both innovation and low-quality replication. But one thing is beyond doubt: half the script of China's LLM era was written inside the Tsinghua campus — not even Stanford or MIT can say the same of their own countries.

Eyewitness Accounts

Call for contributions

If you have studied or worked at Tsinghua University's Department of Computer Science, IIIS, KEG, AIR, or any Tsinghua-affiliated LLM company, please contribute on GitHub.

References

1. 张钹 (Zhang Bo). (2007). 人工智能：现状与未来 [Artificial Intelligence: Present and Future]. Tsinghua University Press.
2. Yao, A. C. (2003). "Classical Physics and the Church-Turing Thesis." Journal of the ACM, 50(1).
3. Tang, J., Zhang, J., Yao, L., Li, J., Zhang, L., & Su, Z. (2008). "ArnetMiner: Extraction and Mining of Academic Social Networks." Proceedings of KDD 2008.
4. Du, Z., Qian, Y., Liu, X., Ding, M., Qiu, J., Yang, Z., & Tang, J. (2022). "GLM: General Language Model Pretraining with Autoregressive Blank Infilling." Proceedings of ACL 2022.
5. Zeng, A., Liu, X., Du, Z., et al. (2023). "GLM-130B: An Open Bilingual Pre-trained Model." Proceedings of ICLR 2023.
6. Hu, S., Tu, Y., Han, X., et al. (2024). "MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies." arXiv:2404.06395.
7. 清华大学计算机科学与技术系 (2018). 清华大学计算机系六十年（1958–2018） [Sixty Years of the Tsinghua CS Department]. Internal publication.
8. Beijing Academy of Artificial Intelligence (2021). "Wu Dao 2.0 Technical Report."
9. Zhipu AI (2023, March 14). "ChatGLM-6B Open Source Release Announcement." GitHub: THUDM/ChatGLM-6B.
10. 财新周刊 Caixin Weekly (2024). "中国大模型六小虎：清华系如何垄断创业第一梯队" [China's Six LLM Tigers: How the Tsinghua Family Cornered the First Tier].
11. CSRankings.org (2025). "AI Subarea Rankings 2020–2025."
12. Shanghai AI Laboratory (2024). "InternLM Technical White Paper."