GLM is the large-model series from Chinese AI company Zhipu (English brand Z.ai, formerly “Zhipu AI”). Zhipu spun out of Tsinghua University lab research in 2019, and most of its founders are Tsinghua professors. Keep one distinction in mind: GLM is the model; Zhipu (Z.ai) is the company.
Zhipu listed on the Hong Kong Stock Exchange on January 8, 2026 (2513.HK), the first of China’s “six AI tigers” to go public. It built its name on open-source GLM. After the US put it on an export-control list, it moved to running its models on Chinese-made chips.
Key Data Snapshot
| Item | Data |
|---|---|
| Models / products | GLM large-model series; also CodeGeeX (coding), CogView / CogVideoX (multimodal), AutoGLM (agent), Zhipu Qingyan (consumer app) |
| Company | Zhipu (Z.ai), headquartered in Beijing, founded 2019 |
| Core team | Founders Tang Jie and Li Juanzi (Tsinghua professors); CEO Zhang Peng |
| Listing | Listed on HKEX 2026-01-08 (2513.HK); IPO raised about US$560 million |
| Flagship model | GLM-5.3 (August 2026), weights released |
| H1 2026 results | Revenue about RMB 954 million (up about 400% YoY); loss for the period about RMB 2.07 billion |
| Developer reach | Over 7.4 million enterprise and developer users on the GLM open platform (2026-08-31, official) |
A Quick Tour of Six Dimensions
① Technology and products: The GLM architecture came from a Tsinghua research team and was published in 2021. Zhipu then ran two tracks: open weights plus a closed-source API. GLM-5.3, released in August 2026, keeps the GLM-5.2 base model; all the gains come from post-training. Its weights came out two weeks later, under Zhipu’s own GLM-5.3 license, not a standard open-source license like MIT. Official benchmark scores are the company’s own figures.
② Customers and market position: In its early years Zhipu mainly sold private deployments to companies and governments. In the first half of 2026, 86.5% of its revenue came from its pay-per-use open platform / API. Its main rivals in China include DeepSeek, Qwen, Kimi, and Doubao.
③ Open-source ecosystem: Model weights are free to download on Hugging Face and similar platforms. The developer community was its first source of reputation and also drives API user growth.
④ Finances: Zhipu raised money twice after listing: a share placement of about HK$31.4 billion (about US$4 billion) in July 2026, and about US$5.02 billion from new shares plus convertible bonds in September. Revenue is growing fast, but R&D spending is still higher than revenue. The reported loss for the first half narrowed from a year earlier, but the “adjusted net loss,” which strips out one-off items, widened.
⑤ Business model and regulation: Chinese AI model makers are in a price war. Zhipu’s gross margin was about 26% in the first half of 2026, still far from profit. To offer generative AI services in China, a company must complete a filing and meet content-safety rules.
⑥ Geopolitics and compute: In January 2025, the US added Zhipu to its export-control Entity List. Zhipu publicly objected, and as of October 2026 it has not been removed. In September 2026, a joint advisory from the US NSA, FBI, and CISA and a report from Anthropic both named Zhipu and other Chinese companies for using “distillation” to take capabilities from US models. These are accusations. On compute, Zhipu’s interim report says its inference runs on a cluster of more than 100,000 Chinese-made chips, and its inference cost per token fell 80% from the start of the year.
Key Milestones
| Date | Milestone |
|---|---|
| 2019 | Zhipu founded (Tsinghua University technology transfer) |
| 2021 | GLM pre-training architecture published |
| 2025-01 | Added to the US Commerce Department’s export-control Entity List |
| 2025-07 | Rebrands externally as Z.ai; releases GLM-4.5 |
| 2026-01-08 | Lists on HKEX (2513.HK) |
| 2026-06 | Releases GLM-5.2 |
| 2026-07 | Share placement of about HK$31.4 billion |
| 2026-08 | Releases GLM-5.3 (8/14) and its weights (8/28); reports H1 results (8/31) |
| 2026-09 | Raises about US$5.02 billion via new shares plus convertible bonds; named for distillation in a US advisory and an Anthropic report |
Table current as of October 2026.
Further Reading
- Other Chinese models: DeepSeek, Qwen, Kimi, Doubao
- How US controls work: Export Controls
- The whole hardware chain: The AI Hardware Supply Chain, End to End
Penchan’s summary
With Zhipu, model scores matter less than two things: whether API revenue can catch up with R&D spending, and whether US controls and the distillation accusations will limit its partnerships outside China.