GLM 5.2: what the model is and how it runs at lowcloud
GLM 5.2 is a language model from the Chinese company Z.ai. The weight files on Hugging Face are dated 16 June 2026. The model has 753 billion parameters, a context window of 1 million tokens, and is published under the MIT license. It handles text only.
At lowcloud, GLM 5.2 is one of 14 models that run on servers in the EU. A request goes to lowcloud UG in Bielefeld, from there to Melious AI GmbH in Saarbruecken, from there to a data centre in the EU. Z.ai never sees the request. You pick the model yourself: per message in chat, fixed for an agent, or for a single step of a workflow. The cost of every run is shown in euros in the app.
- Vendor
- Z.ai, China
- Context window
- 1 Mio. Token
- Reads images
- no
- Licence
- MIT
- Processing
- EU servers via Melious
Where GLM 5.2 comes from
The maker is Z.ai, a company in China. The weight entry on Hugging Face is dated 16 June 2026. The files are publicly downloadable. The license is MIT, one of the most permissive licenses there is. Anyone who downloads the files may use the model commercially, modify it and pass it on, as long as the copyright notice stays with it. For GLM 5.2, the words open source are accurate. For the successor they are not. GLM 5.3 is published on Hugging Face under a Z.ai license of its own with conditions attached, not under MIT. Many overview pages treat both models the same way. That is wrong. The technical facts, briefly. Parameters are the numbers a model learns during training; GLM 5.2 has 753 billion of them. It is an MoE model, short for Mixture of Experts: the model is split into 256 specialist parts plus one fixed part, and only eight of them do the work for each piece of text. Z.ai does not publish a figure for how many parameters are active per piece of text. The context window is 1 million tokens. Tokens are the chunks a model breaks text into, usually parts of words. The context window is the space your instruction, your data and the conversation so far all have to fit into.
What "on EU servers" means here
The chain has three stops: you, then lowcloud UG in Bielefeld, then Melious AI GmbH in Saarbruecken, then a data centre in the EU. Z.ai is not part of that chain. The maker of the model does not see the request. The model runs as a file on European servers. What lowcloud does not know, we say as well. Melious runs no data centres of its own; it distributes load across European operators. Melious does not publish which operator serves GLM 5.2 in any given case. If a client needs a named operator written into a contract, this chain cannot deliver that.
What GLM 5.2 is good for in a workflow
GLM 5.2 takes text and returns text. That covers everything to do with reading, sorting, summarising and drafting. The real reason to pick this model is the context window. A contract of several hundred pages, a full email thread, a tender package made up of several files: all of it fits into one step. You do not have to cut the workflow into pieces and stitch the partial results back together afterwards. The same holds for workflows with many steps in sequence, where the agent still needs to know at step 11 what came out of step 3. One thing worth checking first: the Hugging Face entry lists English and Chinese as the model's languages. German is not listed as a focus. Whether the output in your own language is good enough for text that reaches a client is something to test on a real case rather than assume.
What GLM 5.2 is not good for
GLM 5.2 does not read images. This is not a setting; the model accepts text only. A scan, a screenshot, a photographed delivery note, a PDF with no text layer: for those you need a different model. In lowcloud's EU catalogue these models read images: Kimi K3, Kimi K2.7 Code, Kimi K2.6, GLM 5.3 Flash, Qwen 3.8 27B and Mistral Small 4. GLM 5.2 does not accept fixed JSON schemas. A JSON schema is a template that tells the model which fields the answer must contain, so the next step can process it reliably. Hand one over and an error comes back. That is lowcloud's own observation, tested against the API. It contradicts the vendor documentation: Z.ai lists structured JSON output as a supported feature. The model can write JSON as free text, but the fields are not guaranteed that way. You do not have to move the whole agent because of this. The model can be set per step: GLM 5.2 reads the long document, and a model that accepts schemas turns the result into a clean record. Models in the EU catalogue that accept schemas: Kimi K2.6, GLM 5.1, GLM 5, Qwen 3.5 397B and DeepSeek V4 Flash.
GLM 5.2 or GLM 5.3
GLM 5.2 has been superseded. Its successor GLM 5.3 has been on Hugging Face since late August 2026 and is also in lowcloud's EU catalogue. Both sit on the same base; according to Z.ai the differences come from post-training, not from a new architecture. Take GLM 5.3 for code and for long workflows with many tool calls, meaning steps where the agent calls another application. Z.ai states a 50 percent gain over GLM 5.2 on its own code benchmark. That is a vendor figure from the model documentation, as of September 2026, not an independent test. One difference still speaks for GLM 5.2: in GLM 5.3 the reasoning step before the answer is always on and cannot be switched off. In GLM 5.2 both modes work. For short, repetitive steps in a workflow, running without the reasoning step is faster and shorter. If an agent runs reliably on GLM 5.2 today, there is no reason to switch. If you are building a new agent, GLM 5.3 is the obvious starting point. Neither of them reads images. For that there is GLM 5.3 Flash in the same line.
What GLM 5.2 costs
lowcloud does not publish per-model prices. A margin is added to the purchase price, so any figure named here would be wrong. What can be said: within the EU catalogue, GLM 5.2 sits in the middle on price. It is neither the cheapest nor the most expensive of the 14 models. The actual cost is shown in euros on every single run in the app. After the first test run you know what a case costs and can give your client a real number instead of an estimate. Choosing the model is part of the free account. Without an account you can have a workflow built at agent.lowcloud.de, and lowcloud picks the model for it.
Try GLM 5.2
Picking a model needs a free account. You then choose GLM 5.2 in the chat, per agent, or per step.
Common questions
What is GLM 5.2?
A language model from Z.ai in China, with weight files on Hugging Face dated 16 June 2026. 753 billion parameters, a context window of 1 million tokens, text only, no images. At lowcloud it runs on servers in the EU.
Is GLM 5.2 open source?
Yes. The weights are published under the MIT license on Hugging Face, a genuinely free license with no usage conditions beyond keeping the copyright notice. One caveat for comparisons: this does not apply to the successor. GLM 5.3 is published under a Z.ai license of its own with conditions attached. Overview pages that label both as open source are wrong about 5.3.
Does my data go to China?
No, not in the chain lowcloud runs. The request goes to lowcloud UG in Bielefeld, from there to Melious AI GmbH in Saarbruecken, from there to a data centre in the EU. Z.ai does not receive it. The fact that the model was developed in China does not change that: it runs as a file on European servers. What stays open is which European operator serves it in a given case. Melious does not publish that.
Can GLM 5.2 read images or scanned PDFs?
No. GLM 5.2 takes text only. A PDF with a text layer can be extracted beforehand and passed in; a scan or a photo cannot. For images use one of these models from the EU catalogue: Kimi K3, Kimi K2.7 Code, Kimi K2.6, GLM 5.3 Flash, Qwen 3.8 27B or Mistral Small 4.
Why does GLM 5.2 reject my JSON schema?
Because the provider returns an error for fixed schemas on this model. lowcloud tested this against the API; Z.ai's documentation, by contrast, lists structured JSON output as supported. The model can write JSON as free text, but you do not get guaranteed fields that way. For the step that needs fixed fields, use another model: Kimi K2.6, GLM 5.1, GLM 5, Qwen 3.5 397B or DeepSeek V4 Flash.
GLM 5.2 or GLM 5.3?
For new agents, GLM 5.3, especially for code and long workflows with many tool calls. For existing agents that run reliably, there is no reason to switch. GLM 5.2 has one advantage: the reasoning step before the answer can be switched off, which GLM 5.3 does not allow. For short standard steps that is faster and shorter. Neither reads images.
GLM 5.2 is a product name of Z.ai. There is no business relationship between lowcloud and Z.ai; we run the openly available model through a European provider. Last updated: 7 September 2026.