GLM 5.3: what the model is and how lowcloud runs it on EU servers

GLM 5.3 is a language model from the Chinese vendor Z.ai. The weights have been public on Hugging Face since 25 August 2026. The model has 753 billion parameters, takes text only, and holds a context of roughly 1 million tokens. Tokens are the text fragments a model splits an input into; one token is roughly half a word to a whole word.

At lowcloud you pick GLM 5.3 as the model behind an agent, for a single chat message, or for one individual step of a workflow. The request then does not go to Z.ai. It goes through Melious AI GmbH in Saarbrücken to data centres inside the EU.

Vendor: Z.ai, China
Context window: 1 Mio. Token
Reads images: no
Licence: GLM-5.3-Lizenz
Processing: EU servers via Melious

Where the model comes from

The vendor is Z.ai, based in China. The model repository zai-org/GLM-5.3 on Hugging Face was created on 25 August 2026. Anyone can download the weights and run the model themselves. GLM 5.3 is an MoE model. MoE stands for Mixture of Experts: the model is made of many sub-networks, and only a small part of them computes per token. The configuration file in the repository lists 78 layers and 256 selectable experts plus one that is always used, with 8 of them picked per token. Z.ai does not publish how many of the 753 billion parameters are actually active per token. According to the model card, GLM 5.3 uses the same base model as GLM 5.2. The difference comes from post-training. Z.ai states a 50 percent improvement over GLM 5.2 on an in-house code test set. That is a vendor figure from the model card dated 25 August 2026, not an independent measurement. There is no successor. As of 7 September 2026, GLM 5.3 is the newest model in the series. Alongside it sits GLM 5.3 Flash, a smaller model with a different profile.

The licence is not MIT

Many overview sites list GLM 5.3 as "open source". That is inaccurate. Hugging Face records the licence as "other" with the name "glm-5.3". A separate Z.ai licence applies, neither MIT nor Apache 2.0. In substance it is close to MIT: the copyright notice must travel with any copy, use must comply with applicable law, and there is no warranty. On top of that comes an obligation MIT does not have. Anyone offering the model as a service whose combined revenue with affiliated entities exceeds 10 billion US dollars over any twelve consecutive months must pass a Z.ai security review first. The licence explicitly excludes two cases: products that only embed model capabilities, and plain forwarding of requests to third-party hosted models. For you and for lowcloud this threshold has no practical effect. If your client asks about the licence, the correct answer is still "open weights under a Z.ai licence of its own", not "open source".

What "on EU servers" actually means

The chain is short and nameable: your client, lowcloud UG in Bielefeld, Melious AI GmbH in Saarbrücken, data centres in the EU. Z.ai does not appear in it. The vendor supplies the weights and never sees your requests. Melious runs no data centres of its own. It distributes across European operators and commits to processing staying inside the EU. What lowcloud does not know: which operator serves GLM 5.3 in a given case. Melious does not publish that. We pass on the commitment we have, and nothing beyond it. The difference to the other routes: booking directly with Z.ai means processing in China. Going through a reseller such as OpenRouter means shifting providers depending on load, mostly outside the EU; a fixed assignment per model is not part of that offer. Before you use GLM 5.3 with a client's data, we settle the contractual side with you: a data processing agreement under Art. 28 GDPR and the naming of sub-processors. That belongs in a conversation, not on a web page.

What GLM 5.3 is good for

Long inputs. Roughly 1 million tokens of context is enough for a stack of contracts, a year of ticket history, or a full documentation set, without cutting the material up first. The Z.ai documentation states a maximum output length of 128,000 tokens, so the model can also write long results in one piece. Workflows with many steps. Z.ai aims the post-training of GLM 5.3 at agent work: several steps in a row, with tool calls to search, a mailbox or a CRM in between. That is the case a builder like lowcloud exists for. Text and code work without an image component. Summarising, comparing, rewriting, evaluating text data. Cases where you have to tell your client where processing happens.

What GLM 5.3 is not good for

Images. GLM 5.3 handles text only; the Z.ai documentation says so explicitly, and lowcloud confirmed it against the API. An attached photo or scan is stored and passed along, but not read. If an agent has to evaluate images, use GLM 5.3 Flash, Kimi K2.6 or Mistral Small 4. All three also run on EU servers. Fixed JSON schemas. This is the point where lowcloud's own measurement contradicts the vendor. The Z.ai documentation lists "Structured Output" as supported (retrieved 7 September 2026). On the EU route, a request carrying a fixed JSON schema is rejected with an error. That is lowcloud's own observation from testing the API, not copied from anywhere. Whether the cause sits with Melious or with Z.ai is unresolved. lowcloud catches this and asks for the desired structure in the prompt instead of enforcing it. For a step whose output must match a schema exactly, because another system consumes it directly, a model with schema support is the safer choice: Kimi K2.6, GLM 5.1, Qwen 3.5 397B or DeepSeek V4 Flash, all on EU servers.

What it costs

This page names no price. What a single run cost is shown in euros on every run in the app. That is the number you bill your client from. For orientation: GLM 5.3 sits in the middle of the fourteen EU models. It is neither the cheapest nor the most expensive. You do not have to commit to one model. The choice applies per chat message, fixed per agent, or per individual step of a workflow. That lets you limit an expensive model to the two steps that actually need it.

Try GLM 5.3

Picking a model needs a free account. You then choose GLM 5.3 in the chat, per agent, or per step.

Book a call

Common questions

Does my data go to China if I use GLM 5.3?

No. The request goes from lowcloud in Bielefeld to Melious AI GmbH in Saarbrücken and is processed in data centres inside the EU. Z.ai only publishes the model weights and never sees your requests. Which European operator serves the model in a given case is not published by Melious.

Is GLM 5.3 open source?

The weights are public and freely downloadable, but the licence is neither MIT nor Apache. A separate Z.ai licence applies. It requires the copyright notice to travel with copies and, for providers with more than 10 billion US dollars of revenue over twelve months, a prior security review by Z.ai. For your use it has no effect.

Can GLM 5.3 read images?

No, the model handles text only. For photos, scans or screenshots use GLM 5.3 Flash, Kimi K2.6 or Mistral Small 4. All three are also processed on EU servers.

How big is GLM 5.3?

753 billion parameters in total. It is an MoE model, so only part of that computes per token: 8 of 256 selectable experts plus one always-on expert, across 78 layers. Z.ai does not state how many parameters that amounts to per token.

What is the difference between GLM 5.3 and GLM 5.3 Flash?

Both hold roughly 1 million tokens of context, but GLM 5.3 Flash also reads images and is the cheaper of the two. GLM 5.3 is the larger model and is aimed at long workflows with many tool calls. Both are available on EU servers at lowcloud.

What do I need to use GLM 5.3 at lowcloud?

A free account. Without an account you can have a workflow built for you at agent.lowcloud.de, and lowcloud picks the model. Choosing the model yourself comes with the account.

GLM 5.3 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.