# Z.ai (Zhipu AI)

Z.ai, also known as Zhipu AI (智谱), is the Chinese lab that makes the GLM models. Z.ai is also the name of its overseas platform. The domestic one is BigModel.cn. According to Zixuan Li, its director of product, in the November 2025 ChinaTalk interview summarized on [[zai-playbook]], Z.ai's overseas API is hosted in Singapore and run by a Singaporean company that employs him, with inference kept outside China to meet data-storage requirements for foreign customers. Training happens in China. In Nathan Lambert's July 2026 ranking of labs by their best model, Z.ai was fifth with GLM-5.2, behind Anthropic, OpenAI, Moonshot and SpaceXAI ([[kimi-k3-open-weights-escalation]]). By August its GLM-5.3 beat Kimi K3 on many benchmarks.

## History

Zhipu was founded in 2019, and at first it was not working on language models. Li says it built AMiner, an academic search and scholar-mapping database similar to Google Scholar and based on graph methods. It moved to large language models in 2020 and published the GLM paper in 2021. Lambert's [[glm-5-3-how-chinese-labs-keep-stride]] credits that first model to THUDM, Tsinghua University's Data Mining and Knowledge Engineering group, and Li puts it about a year before GPT-3.5. The lab has kept close ties to Tsinghua, and Li describes many PhD students working on GLM while doing their degrees.

| Model | Date | Notes |
|-------|------|-------|
| GLM | March 2021 | THUDM paper, weights published |
| GLM-130B | August 2022 | scaled version |
| ChatGLM, ChatGLM2, ChatGLM3 | March, June, October 2023 | chat models, 6B weights published |
| GLM-4 | January 2024 | rebrand to "GLM"; flagship closed, GLM-4-9B weights in June |
| GLM-4.5, GLM-4.6 | 2025 | 355B; the releases that made Z.ai known outside China |
| GLM-5 | February 2026 | next major generation |
| GLM-5.2 | June 2026 | MIT-licensed weights; [[glm-5-2-step-change-for-open-agents]] |
| GLM-5.3 | August 2026 | same base as 5.2, ~750B, extended post-training |

The first six rows follow Lambert's timeline in the GLM-5.3 post. The 355B figure is Li's. Other pages in the vault fill gaps: [[reap-expert-pruning]] gives GLM-5.1 as 744B total and 40B active, and [[qwen3-8-flash-next]] records GLM-5.3-Flash at 320B total and 18B active. The routing shapes of GLM-5.2 (256 experts, 8 active per token) are in [[glm52-amd-mi355x]], a vendor account of quantizing the model to MXFP4 and serving it on AMD hardware.

Li says the lab became famous with GLM-4.5 and 4.6 because they were good at coding, reasoning and agentic tool use. Lambert called GLM-4.5 "another DeepSeek moment". Li traces the change in focus to 2025, when Manus and [[claude-code]] convinced the company that coding and agent work was worth more than chat. It had previously optimized for Chatbot Arena, where it ranked roughly sixth to ninth. GLM-4.5 was built by training three separate teacher models for reasoning, agents and coding, then distilling them into one. In August 2026 Z.ai described GLM-5.3 in one sentence: "Scaling post-training is all we did." Lambert says Z.ai's strength is post-training, while Moonshot's is pretraining.

## Why it releases weights

Li gives two reasons. The first is contributing to research, as Llama, Qwen and Kimi do. The second is commercial. American users will not send data to a Chinese company's API, so a Chinese model has to be available on Fireworks, Groq or the customer's own hardware to get used in the US. Z.ai's flagship was closed in 2024, and DeepSeek R1 changed its mind by showing a lab could become famous from open weights and still earn money through APIs and partnerships. Li's line for this is "You need to expand the cake first and then take a bite of it." Lambert's Kimi K3 essay gives the same practical explanation for Chinese labs in general.

Li also describes something that surprised his interviewers. Chinese tech media covers the Chinese models that American influencers talk about, so recognition in Silicon Valley helps sales at home. He named Karpathy, Altman and Musk as the voices that matter. At the time Z.ai had 20,000 followers on X to DeepSeek's million, and Li still saw Reddit users asking what GLM was. By June 2026 Lambert was describing Moonshot and Z.ai as having taken the top of the "reputational market" among AI researchers.

## Business

Releasing flagship weights means customers ask why they should pay Z.ai for a model they could host themselves. Li's answer in 2025 was better engineering: faster decoding, built-in search and MCP tools, and the GLM Coding Plan subscription, which he says keeps users because they stop worrying about how many tokens each prompt uses. The US brought in half of Z.ai's overseas revenue even though India had more users. Li also names role-play as a market, with GLM used on SillyTavern and Janitor AI. By July 2026 Bloomberg reported, as cited by Lambert, that Z.ai was set to be the first Chinese AI company with $1B in annual sales, largely from on-premises deployments. Lambert also says Artificial Analysis scores affect Z.ai's stock price, which is one reason he thinks it cares a bit more about public benchmarks than American labs do.

Its distribution strategy shows up in the vault's toolbox. Li said "we need to combine with these top products to gain fame" and named Claude Code and Kilo Code. Z.ai sponsors [[claude-code-router]], whose README carries a GLM Coding Plan banner, and [[cc-mirror]], which lists Z.ai as a provider. Lambert used GLM-5.2 through Fireworks inside Claude Code. Hosted GLM builds appear on [[telnyx]] (as GLM-5.2-FP8) and [[toolbox/crofai|CrofAI]]. [[mostaque-internet-offline]] reports that Hugging Face used GLM to defend itself during the OpenAI-agent breach, because the restricted American cyber models were not available to it.

## How its release practice changed

Z.ai's release practice changed a lot within a year. In November 2025, asked how long it takes to release a finished model, Li said: "Get it out fast. We open source it within a few hours." Z.ai did not give LM Arena or evaluation companies early access, did not tease models before launch, and treated the open-source release itself as the main event. Li said he personally wanted a week to coordinate with inference providers and coding-agent companies, and instead often had a few hours to call partners in the middle of their night. Lambert quotes the line in his reading list as evidence of how fast Chinese labs move.

GLM-5.2 in June 2026 took three days. It went to Coding Plan subscribers on a Saturday, and the weights and blog followed on June 16. GLM-5.3 in August took longer still. It launched in the coding plan only, with the API promised soon and the weights promised in two weeks. Z.ai then announced a staged security release: selected security partners would evaluate the model first, and full weights would follow once safety evaluations were complete. It described a request classifier and chain-of-thought monitoring on its own platform. By American standards this is still fast, and Lambert considers release speed the main reason Chinese labs keep pace. It is also the first time the vault's sources show Z.ai putting a stated dual-use check between a finished model and its weights. The sources do not say whether the GLM-5.3 weights were released on schedule.

## Scrutiny

Z.ai is not named in any of the congressional letters collected on [[us-scrutiny-of-chinese-model-use]]. Those probes focused on companies using Kimi and Qwen. The political concern Lambert raises in the GLM-5.2 essay does apply to it, though. A Chinese lab releasing more cyber-capable open models while the US restricts its own frontier models invites a US declaration that a specific Chinese model is unsafe. GLM-5.3's staged release is the first Z.ai release where the company itself acknowledged that risk.