GLM is the family of large language models built by Zhipu AI, a Beijing-based lab that rebranded internationally as Z.ai. Unlike ChatGPT or Claude, whose underlying model weights stay locked inside their makers’ servers, most GLM models are released as open weights: anyone can download them, run them on their own hardware, fine-tune them, and use them commercially without a licensing fee. The newest flagship, GLM-5.2, drew attention in mid-2026 for matching leading US models on coding and reasoning benchmarks at roughly a sixth of the API price.

Who makes GLM

Zhipu AI was founded in 2019 as a spinoff from Tsinghua University’s Knowledge Engineering Group, one of China’s leading AI research labs. It bet early on large language models, before the ChatGPT boom made that the obvious strategy. In January 2026 the company went public on the Hong Kong Stock Exchange at a market capitalization of roughly $6.6 billion — the first major AI model developer anywhere to complete an IPO. Internationally, the company now markets itself as Z.ai, while GLM remains the name of its model line.

What the GLM family looks like

GLM stands for General Language Model. Zhipu has iterated through several generations — GLM-4.5, GLM-4.7, and the GLM-5 series — each released as downloadable weights under a permissive MIT license, alongside a hosted API for people who’d rather not run the model themselves. The current flagship, GLM-5.2, released in June 2026, is a mixture-of-experts model: roughly 753 billion parameters in total, but only about 40 billion are active for any given piece of text, which keeps inference cheaper than a dense model of the same size. A long-context variant extends its window to one million tokens — enough to feed in a large codebase or a lengthy document collection in a single request.

What made GLM-5.2 notable

Two things stood out about GLM-5.2’s launch. First, independent trackers found it competitive with GPT-5.5-class models on coding and agentic tool-use benchmarks, while Z.ai’s API charged roughly a sixth of the price — a gap large enough that developers building AI coding assistants took notice. Second, Zhipu said the model was trained entirely on Huawei’s Ascend chips, with no Nvidia hardware involved at any stage. That matters because US export controls restrict Chinese firms’ access to Nvidia’s most advanced AI chips — GLM-5.2 became a visible data point in the argument over whether those controls are actually slowing Chinese frontier-model progress.

What “open weights” means in practice

Open weights don’t mean the model was built in the open, the way open-source code is — Zhipu hasn’t published full training data or methodology. It means the finished, trained model file is public: a developer, company, or researcher can download it from a hub like Hugging Face and run it on their own servers, modify it, or build a product on top of it, all under the MIT license’s permissive terms. That’s the same playbook other Chinese labs like Alibaba’s Qwen and DeepSeek have used to win developer adoption — compete on openness and price rather than trying to out-spend US labs on proprietary training runs. For a broader look at what running a model this way involves, see our explainer on open-weight AI models and on large language models generally.

The actual cost

Self-hosting GLM-5.2 is free of licensing fees, but you still pay for the computing hardware to run a 753-billion-parameter model, which is out of reach for most individuals. The simpler path is Z.ai’s hosted API or its free web chat at chat.z.ai. As of July 2026, per Z.ai’s official pricing page, the GLM-5.2 API costs $1.40 per million input tokens and $4.40 per million output tokens, with cached input discounted to $0.26 per million — Z.ai also offers smaller “Flash” models in the family free of charge. Prices for AI APIs change often, so check the pricing page for the current rate before committing to a project.

In the news

GLM-5.2’s price and performance made headlines when it shipped: see our report on how GLM-5.2 rivals US frontier AI at a sixth of the cost.

FAQ

Is GLM the same as ChatGPT? No. ChatGPT is built on OpenAI’s proprietary GPT models; GLM is a separate model family from the Chinese lab Zhipu AI (Z.ai), and — unlike GPT — most GLM versions are openly downloadable.

Can I legally use GLM in a commercial product? Generally yes — Zhipu releases most GLM models under the MIT license, which permits commercial use, modification, and redistribution. Always check the specific license attached to the version you download, since terms can vary by release.

Is GLM free to use? The web chat and some smaller “Flash” models are free. Running the full GLM-5.2 model yourself costs computing resources, and Z.ai’s hosted API bills per token, though at a lower rate than comparable US models.

Why does GLM-5.2 run on Huawei chips instead of Nvidia’s? Because US export controls limit Chinese companies’ access to Nvidia’s most advanced AI accelerators, so Zhipu trained GLM-5.2 entirely on domestic Huawei Ascend hardware instead.