Aiodic

Back to the table

70 Model family Gl GLM Model Families

GLM

Zhipu AI's model family.

What it is

GLM is a family of AI models made by Zhipu AI, a Beijing company that has called itself Z.ai outside China since July 2025. Like GPT, Claude, and Qwen, the GLM models are language models: systems trained on huge amounts of text to predict and write words, so they can answer questions, draft text, translate, and write code. The name stands for General Language Model.

GLM is best known for two things. The first is that Zhipu has published many of its models as "open weights", meaning anyone can download the finished model files and run them on their own hardware. The second is coding: since mid-2025 the company has aimed its flagship models squarely at coding agents, the tools that write, run, and fix code with little help, and it sells a subscription built for them.

Where it came from

Zhipu was founded in 2019 as a spin-out from Tsinghua University, one of China's leading universities, by researchers from its Knowledge Engineering Group, including the computer science professors Tang Jie and Li Juanzi. In October 2022 the Tsinghua team published GLM-130B, a model with 130 billion parameters (the adjustable numbers a model learns during training) that understood both English and Chinese. The paper, later accepted at the ICLR 2023 conference, claimed it beat OpenAI's GPT-3 on a range of English tests and could run on four ordinary gaming graphics cards once compressed. Its weights were shared with researchers, which was unusual for a model that large at the time.

On 14 March 2023, a few months after ChatGPT arrived, Tsinghua and Zhipu released ChatGLM-6B, a 6.2-billion-parameter chat model trained on about a trillion tokens of Chinese and English text. Compressed, it needed only about 6 GB of graphics memory, so hobbyists could run it at home. It became one of the most popular early open models in China, and the team later reported more than 10 million downloads of its open models on Hugging Face, the main website for sharing models, in 2023 alone.

The generations, in brief

Model families change quickly, so treat this as a snapshot at the time of writing. Dates come from Zhipu's own release notes where possible.

Zhipu also makes image, video, voice, and text-recognition models, and AutoGLM, an agent that operates a smartphone's apps for you.

Open weights, not open source, and licences that vary

"Open weights" is not the same as "open source" in the full sense. Zhipu publishes finished models and technical reports, but not its complete training data, so you can use and adapt the models but cannot rebuild them from scratch. The licences have also changed several times, which is easy to miss:

Always read the licence file of the exact model you download. The Parameters page explains which sizes of model will run on which kinds of computer.

Why it matters

It is one of the main sources of strong open models. Along with DeepSeek, Qwen, and Moonshot AI's Kimi, GLM is part of the wave of Chinese open-weight models described on those pages. When GLM-4.5 launched, Z.ai said it ranked third overall, and first among open models, across 12 tests it had chosen. Claims like this come from the company; check independent benchmarks and your own tests.

It is aimed at coding agents. Z.ai has built its recent models for long, multi-step programming work and sells them through a monthly plan designed to plug into other companies' coding tools. That makes GLM one of the ways people run a coding agent at a lower price than the best-known Western models, a trade-off the examples below explore.

It is also a political story. Zhipu is on a US government trade blacklist, is backed partly by Chinese state funds, and is subject to China's content rules. Where your data goes and what the model will say depend heavily on how you reach it, which matters as much as any benchmark.

Real-world examples

The US Entity List and OpenAI's warning (January to June 2025)

On 15 January 2025, in the last days of the Biden administration, the US Commerce Department's Bureau of Industry and Security added Beijing Zhipu Huazhang Technology and nine related Zhipu companies to its Entity List, a trade blacklist. The notice, published in the Federal Register the next day, said they were added because they advance China's military modernisation "through the development and integration of advanced artificial intelligence research". In practice, American companies need a licence, which will normally be refused, to supply them with anything covered by US export rules, including advanced chips. The South China Morning Post reported that Zhipu said the decision "lacks a factual basis" and would not have a substantial impact on its business.

On 25 June 2025, OpenAI's global affairs team published a post naming Zhipu as a company making notable progress in getting governments abroad to adopt Chinese AI. It said Zhipu offered "sovereign" AI systems to governments in Southeast Asia, the Middle East, and Africa, including hardware built with Huawei, and put state-backed investment in the company at over 1.4 billion US dollars. These were OpenAI's claims about a direct competitor, published during a policy argument in Washington, so they are one side's account. They do show why GLM's story involves governments as much as software.

Being blacklisted also pushed Zhipu towards Chinese chips. In January 2026 the company said its GLM-Image model had been trained entirely on Huawei's Ascend chips. Network World noted that Zhipu did not say how many chips or how long it took, so the claim is hard to check independently.

GLM-4.5, the Coding Plan, and a sign-up limit (July 2025 to January 2026)

On 28 July 2025 Zhipu released GLM-4.5 and GLM-4.5-Air under the MIT licence and presented itself internationally as Z.ai. Computerworld reported that the chief executive, Zhang Peng, told CNBC the model needed just eight Nvidia H20 chips to run. Z.ai described the model as "agent-native", built to reason, write code, and use tools in one model, and made it work with Anthropic's Claude Code and other coding tools, so developers could swap it in behind a tool they already used.

Z.ai sells this as the GLM Coding Plan, a monthly subscription for use inside coding agents rather than a chat window. At the time of writing its documentation lists Lite, Pro, and Max tiers starting at 18 US dollars a month, with usage allowances that reset every five hours and weekly, and names Claude Code, Cline, and OpenCode among the supported tools. Prices, tiers, and limits have changed more than once, so check the current page.

Demand brought problems. After GLM-4.7 launched in December 2025, users reported slow responses and rate limits at busy times, and on 21 January 2026 Zhipu cut new sign-ups for the Coding Plan to 20 percent of their previous daily level, the South China Morning Post reported, blaming a shortage of computing power and saying it would crack down on "malicious" use. Existing subscribers kept their plans. The episode is a useful reminder that a cheap service is only useful if it has the capacity to answer.

China's first listed language-model company (January 2026)

On 8 January 2026 Zhipu listed on the Hong Kong Stock Exchange, under stock code 2513, after raising about 558 million US dollars. Its own announcement called it "the world's first listed large language model company" and claimed more than 12,000 business customers and 45 million developers as of September 2025. Those are the company's figures.

The listing also put its finances in public view. Rest of World reported that Zhipu lost roughly 330 million US dollars on about 27 million dollars of revenue in the first half of 2025, and that most of its revenue came from Chinese state-owned enterprises and financial institutions. In its first annual results, reported by the South China Morning Post on 31 March 2026, revenue rose 132 percent to 724 million yuan (about 105 million US dollars) for 2025, slightly below analysts' estimates, while total losses grew to 4.72 billion yuan. Heavy spending ahead of revenue is common among AI labs, and it is worth knowing when you depend on one.

A US government test of GLM-5.3 (September 2026)

GLM-5.3, released in mid-August 2026, was notable for something Z.ai itself flagged: its skill at finding and exploiting software security flaws had grown faster than the company expected. On 17 September 2026 the US Center for AI Standards and Innovation (CAISI), part of NIST, published its own assessment. It found GLM-5.3 was "the most cyber-capable open-weight model released to date", but that its abilities were "significantly lower" than those of current US frontier models, lagging them by about four months on its combined measure. CAISI noted it could not compare models that had been built but not yet released. This is a US agency testing a Chinese rival's model, so read it as one careful source rather than the final word, but it is a rare independent check on a model maker's claims.

An everyday case: the freelancer's coding budget

Here is an illustrative scenario, not a real person. Priya builds websites for small businesses and uses a coding agent every day. Her usual subscription keeps running out mid-week, so she tries pointing the same tool at a GLM model through Z.ai's plan. For routine tasks, such as adding a contact form or fixing a layout bug, she finds the results close enough, and it costs less. For a client in healthcare she stops: the code and data would go to servers run by a Chinese company, and the client's contract doesn't allow that. For that project she goes back to her usual provider. She keeps a note of which clients allow which services, and still reviews every change the agent makes.

Try it yourself

This exercise lets you run a GLM model on your own computer and see how licences differ between versions. You don't need an account. It uses Ollama, a free program for running open models locally; the Parameters page explains more about model sizes and memory.

Hardware at the time of writing: Ollama's glm4 model (GLM-4-9B, from 2024) is a 5.5 GB download, and a computer with 16 GB of memory is a comfortable starting point. The newer glm-4.7-flash is a 19 GB download and needs a much more powerful machine. Sizes change, so check the Ollama pages first.

  1. Install Ollama. Download it from ollama.com for Windows, Mac, or Linux and install it like any other program.
  2. Open a terminal. On Windows, open the Start menu and type "PowerShell". On a Mac, open the "Terminal" app.
  3. Download and start the model. Type this and press Enter. The first time, it downloads the model: > ollama run glm4:9b
  4. Give it a bilingual task. GLM was built for Chinese and English, so test that. Paste: > Write a short, polite email in English to a supplier asking to move a delivery from Tuesday to Thursday. Then translate it into Simplified Chinese, and list three phrases in the translation that a native speaker might word differently. If you know someone who reads Chinese, ask them to judge the result. Notice that the model is grading its own translation, which is not the same as checking it.
  5. Try a coding task. Paste: > Write a simple HTML page with a button that counts how many times it has been clicked. Explain each part in plain English. Save the code as a file ending in .html and open it in your browser. Does it work first time?
  6. Close it by typing /bye and pressing Enter.
  7. Read the licence. Go to huggingface.co/zai-org (the resources below link to it) and find a recent model such as GLM-5.2, then open its "LICENSE" file. Compare it with the GLM-4-9B licence summarised on this page. Which would let a small business use it with no paperwork? This is the habit that matters most with open models.
  8. Optional: compare where you run it. Ask your local model and Z.ai's own website, chat.z.ai, the same neutral question, then the same question about recent Chinese history. In a single test published in July 2026, Bharath Reddy of the Takshashila Institution found that GLM-5.2 on Z.ai's own chat returned a blank answer or a warning on sensitive questions, while the same model through a third-party host answered in detail. Anything you type on chat.z.ai goes to Z.ai's servers, so keep it harmless.

Common mistakes and misunderstandings

Thinking "GLM" is one model. It is a family spanning five years, from 6-billion-parameter chat models to ones with over 700 billion, plus vision, voice, image, and phone-agent models. Always note the exact name and version.

Assuming every GLM model is MIT-licensed. Many recent ones are, but older models require registration for commercial use and are governed by Chinese law, GLM-5.3 has a custom licence with a security review for very large companies, and some versions are closed. Read the licence of the exact file you download.

Confusing Zhipu, Z.ai, ChatGLM, and GLM. Zhipu is the company, Z.ai is its international brand and website, GLM is the model family, and ChatGLM was both an early model name and the Chinese consumer app. News stories use all four.

Assuming censorship is only in the website. The July 2026 Takshashila test suggested much of the filtering happens in Z.ai's own service. But a March 2026 study by Gregory Frank, run on downloaded weights, found that GLM-Z1 answered all of its English questions on Chinese politics accurately but refused some of the same questions asked in Chinese. So some behaviour is trained into the model itself, and it can depend on the language you use. The DeepSeek and Qwen pages cover the wider picture; Western models have their own biases too.

Forgetting where your data goes. Downloading a model and running it on your own machine keeps your data there. Using Z.ai's website, app, API, or Coding Plan sends it to a Chinese company. Using a third-party host means that company's terms apply. The Privacy page covers what to check, and some employers forbid services from companies on the US Entity List.

Treating benchmark claims as settled. Z.ai, like every model maker, publishes scores from tests it chose, and headline rankings change within weeks. Test on your own tasks, as the Evals page explains.

Expecting it to stay the same. Z.ai has released a new flagship roughly every two to four months since mid-2025, and older API model names are redirected to newer ones. Anything here about versions, prices, or terms may be out of date by the time you read it.

Related elements