Tsinghua University

Contribution to AI

Two things had to exist before China could have a large-model industry of its own: researchers trained on the right problems, and weights somebody was willing to publish. Both came out of the same corridor of the computer science department. Zhang Bo, who had worked on artificial intelligence at the university since the late 1970s, argued that neither symbolic reasoning nor deep learning would suffice alone and that a third generation of the field would have to fuse the two — a framing that shaped how Chinese funding bodies subsequently described the problem, and the basis on which the Institute for Artificial Intelligence was assembled in 2018 with Andrew Yao chairing its academic committee. The Knowledge Engineering Group's line ran from AMiner, an academic search and knowledge-graph system operating since 2006, to the GLM objective, which trains by infilling deleted spans rather than predicting strictly left to right. GLM-130B, released in August 2022, was distributed at a scale where Chinese models had previously been announced rather than handed over, and was engineered to run inference on a single eight-GPU node; ChatGLM-6B followed in March 2023 and ran on hardware a student could afford. The natural language processing group's parallel effort produced OpenBMB and the MiniCPM models sized for phones. The staffing pattern is the deeper contribution: Zhipu, Moonshot and ModelBest were founded by Tsinghua professors or their former students.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).