Last updated: 2026-07-03T16:29:41.010Z

From Tsinghua Lab to Hong Kong's Trading Floor: The Z.ai Story

Founded by two Tsinghua professors out of a university research lab in 2019, Z.ai (formerly Zhipu AI) grew into one of China's "AI tiger" companies and became the first large-language-model developer in the world to go public, listing in Hong Kong in January 2026.

Z.ai: From a Tsinghua Research Lab to the World's First Public LLM Company

In 2019, two Tsinghua University professors — Tang Jie and Li Juanzi, both from the university's Department of Computer Science and Technology — turned years of academic research into a company. Zhipu AI, later rebranded internationally as Z.ai, grew directly out of Tsinghua's Knowledge Engineering Group (KEG), a lab the pair had built around knowledge graphs, natural language processing, and social network analysis.

The early days were far from glamorous. As one of many academic spin-offs without an obvious business model, the founders initially struggled to attract investors. It was the administrative commission of Beijing's Zhongguancun Science Park that gave the fledgling team a real start, offering rent-free office space for three months at the Tsinghua University Science Park — often nicknamed China's "Center of the Universe" for tech startups.

Betting on Large Models Early

The decision that would define the company's future came in 2020 — a full two years before ChatGPT made large language models a household concept. While Zhipu continued running its original knowledge-graph business, Tang Jie and his team began quietly investing in foundation-model research, a move that set them apart from most Chinese AI companies of the era, which were still focused on computer vision or speech recognition.

That early bet paid dividends. In 2021, the team published CogView, a 4-billion-parameter Transformer-based text-to-image model presented at NeurIPS, using a VQ-VAE tokenizer to compete directly with OpenAI's DALL-E. A year later, in August 2022, Zhipu and Tsinghua's KEG released GLM-130B, a bilingual English-Chinese pretrained model with 130 billion parameters — one of the earliest large-scale foundation models to come out of China.

Going Open, Going Public with Developers

The real inflection point came in 2023. Zhipu released ChatGLM-6B and the larger ChatGLM-130B as open-weight models, and the Chinese-language developer community embraced them almost immediately. ChatGLM became one of the most widely used open-source Chinese language models of the year, spawning a large ecosystem of fine-tuned derivatives and cementing Zhipu's reputation as one of the "AI tiger" companies — a group of well-funded Chinese large-model startups that also came to include Moonshot AI, Baichuan, and MiniMax.

The pace didn't slow. Zhipu also released CodeGeeX, a multilingual code-generation model trained on more than 850 billion tokens across 20-plus programming languages, distributed as a free plugin for Visual Studio Code and JetBrains IDEs. Then, in January 2024, the company unveiled GLM-4, its next-generation flagship. GLM-4 introduced a 128,000-token context window, native multimodal understanding across text, image, and video, and an "All Tools" agentic feature that let the model autonomously use a web browser, run code, and generate images. Internally, Zhipu benchmarked GLM-4 as competitive with GPT-4 on Chinese reasoning tasks, English MMLU, GSM8K, and HumanEval.

By March 2024, CEO Zhang Peng — also a Tsinghua-trained engineer who spent more than two decades in academia before commercializing his research — announced the company was building a Sora-style text-to-video system. It launched in July 2024 as "Ying" (Qingying), capable of generating six-second clips, and was later open-sourced as CogVideoX, a diffusion-transformer model built around a 3D causal variational autoencoder.

Headwinds and a Rebrand

2025 delivered a mix of pressure and momentum. In January, the U.S. Commerce Department added the company to its Entity List, citing national security concerns — a designation that has shaped how the company sources compute ever since, pushing it toward domestic chips from Huawei and Cambricon rather than Nvidia. Just months later, in April, Zhipu disclosed it had begun preparing for an IPO and released two more models under the permissive MIT license. In May, it signed a 61.28-million-yuan government deal for city projects in Hangzhou.

Then, in July 2025, the company rebranded internationally as Z.ai alongside the release of GLM-4.5 and GLM-4.5 Air — while committing to release its GLM models as free, open-source software under the MIT License going forward.

A Historic Listing

Z.ai's defining moment arrived on January 8, 2026, when it completed its Hong Kong IPO under the legal name Knowledge Atlas Technology Joint Stock Co., Ltd. (HKEX: 2513) — becoming the first large-language-model developer in the world to go public. The offering priced at HK$116.20 a share, raised approximately HK$4.17 billion, and was retail-oversubscribed more than a thousand times over, valuing the company at roughly US$6.5–7 billion. Around 70% of the proceeds were earmarked for continued model R&D through 2028.

The company's debut annual results told a familiar story for China's AI sector: 2025 revenue of about RMB 724 million (roughly $105 million), up 132% year-on-year, against a widened net loss of approximately RMB 4.7 billion — heavy losses driven largely by enterprise-focused, on-premises deployments to state-owned enterprises and financial institutions.

Where Z.ai Stands Today

Since listing, Z.ai has kept shipping at a rapid clip. GLM-5 launched in February 2026 as an open-weight flagship reportedly trained without Nvidia chips, followed by GLM-5.1 in April with a sharper focus on long-horizon agentic coding. By June 2026, GLM-5.2 extended context length to one million tokens and topped several open-weight coding benchmarks by the company's own testing — arriving just days after U.S. export controls temporarily restricted access to Anthropic's Claude Fable 5, a timing Z.ai's leadership was quick to highlight as evidence for the resilience of open-weight models.

From a rent-free office in Zhongguancun to a Hong Kong trading floor, Z.ai's trajectory reflects a broader pattern reshaping the global AI landscape: elite Chinese university labs increasingly spinning out into serious, well-funded competitors on the world stage — one still anchored, by its founders' own account, in the long-term pursuit of artificial general intelligence rather than short-term profit alone.

Explore Z.ai on Nexorzo

Want to try Z.ai's GLM models yourself? Check out the full Z.ai listing on Nexorzo for features, pricing, and use cases.