# From Qinghua lab to Hong Kong listing: Zipu’s path into China’s AI industry

By Riza Dagoc

Canonical URL: https://www.tokenpost.com/news/ai/46181
Published: 2026-11-02T06:20:51.000Z
Updated: 2026-11-02T06:20:51.000Z
Section: AI

> Asianometry traces Zipu’s rise from Qinghua University research to a Hong Kong listing, examining how commercialization and the GLM model series built its technical foundation.

Zipu’s rise from a university research project to a Hong Kong-listed AI company rested on commercializing academic work and developing its own GLM language models, according to Asianometry’s account of the company’s history.

In a video published on November 1, Asianometry operator Jon Y traced the company’s origins at Qinghua University and its development of foundation models. Zipu, also known as ZAI, was one of two AI labs that listed on the Hong Kong Stock Exchange in early 2026, he said.

The business grew out of A Minor, a research intelligence service developed by the university’s Knowledge Engineering Group. It extracted researcher information to build a searchable knowledge graph and rankings. Jon Y said the team began earning revenue from private-sector partnerships in 2015, reflecting a laboratory culture of turning papers into products. Following negotiations over commercialization, Zipu was formally established in June 2019, with the university transferring the service to the company and retaining a significant shareholding.

The release of GPT-3 in 2020 prompted Zhang Peng to pursue large language models, Jon Y said. Its ability to generate text without the elaborate sentence-processing pipelines used in traditional natural language processing represented a breakthrough. With support from his colleagues, Zhang pushed ahead with a bilingual model containing billions of parameters.

Co-founder Tang Jie led Wudao, a government-backed effort involving 100 scientists from several Chinese universities, according to the account. The project produced the GLM framework, which combined elements of GPT and BERT: it used the context surrounding a blank while generating the missing text one token at a time.

Training GLM-130B in 2022 involved hardware failures, unstable model gradients and unexpected memory demands, Jon Y said. The successful training run began in May on 96 Nvidia A100 servers and finished in July. Compute costs were about 4 million RMB, while total costs reached about 10 million RMB. The team aimed to produce a model that could run on a single A100 server.

The decision to release the model openly reflected both limited prospects for immediate revenue from keeping it closed and a research culture that encouraged sharing, Jon Y said. In November 2022, GLM-130B was the only Chinese model included in a Stanford University assessment of 30 models from 12 providers. The assessment acknowledged progress in open models, although GLM-130B lagged closed models and investor interest remained muted.

ChatGPT’s release on November 30, 2022, changed that reception, according to Jon Y. Research that had attracted little public attention suddenly drew widespread interest, bringing Zipu greater visibility as one of the Chinese companies with its own large language models. The company subsequently worked through the Lunar New Year holiday to complete its chatbot and prepared the smaller ChatGLM6B for open release.