Qwen 3.8 Max
Qwen 3.8 Max is the top tier of the Qwen 3.8 model series. It was built by Qwen, the AI division of the Chinese company Alibaba. It is a text model: it reads text and writes text. What sets it apart from most other models in the catalogue is that it works through its reasoning in writing before every answer, and that step cannot be switched off.
On lowcloud, Qwen 3.8 Max is one of 14 models that run on servers in the EU. You pick it in chat for a single message, fix it for one agent, or set it for a single step of a workflow. Processing does not go to Alibaba. It goes through Melious AI GmbH in Saarbrücken to data centres in the EU. Model choice comes with the free account. Without an account, agent.lowcloud.de will still build a workflow for you, but it picks the model itself.
- Vendor
- Qwen, China
- Context window
- 256k Token
- Reads images
- no
- Licence
- Qwen3.8-Max-Lizenz
- Processing
- EU servers via Melious
Where Qwen 3.8 Max comes from
The maker is Qwen, the AI division of Alibaba, based in China. The model belongs to the Qwen 3.8 series, which came out in August 2026. Qwen does not state a separate release date for the Max tier. Alibaba's documentation for it was last updated on 2 September 2026 and lists a revised version named qwen3.8-max-0902. The weights of Qwen 3.8 Max are not published. Weights are the trained numbers a model consists of; without them nobody can run the model themselves. Qwen 3.8 Max is usable only through an interface at a provider that operates it. A related model is public, named Qwen3.8-2.4T-A95B. Qwen describes Qwen 3.8 Max as the managed version of it. That open model has 2.4 trillion parameters, of which 95 billion are active per request. This design is called Mixture of Experts, or MoE: the network is made up of many sub-networks and switches on only a small share of them per request. Qwen publishes no separate parameter count for the Max tier. The licence of the open model is neither MIT nor Apache 2.0. It is called the Qwen3.8-Max License, it comes from Qwen itself, and it carries conditions: products with more than 100 million monthly active users or more than 20 million US dollars in monthly revenue must show the model name visibly in their interface. Anyone reselling model access above 50 million US dollars in annual revenue needs separate permission from Qwen. Overview pages that simply write "open source" here are inaccurate. As of 7 September 2026, Qwen 3.8 Max has not been replaced. No successor at the Max tier is findable; the newer Qwen releases sit below it.
What "on EU servers" actually means here
The chain has three stops. You work in lowcloud. lowcloud, that is lowcloud UG in Bielefeld, passes the request to Melious AI GmbH in Saarbrücken. Melious runs it in data centres in the EU. The request does not go to Alibaba. The maker of the model is not involved in running it. There is one thing lowcloud does not know, and says so: Melious operates no data centres of its own but distributes the load across European operators. Which operator serves a given model is not published by Melious. If your client asks for the specific operator, the answer you can stand behind is: EU, through a named German intermediary, operator not disclosed.
What lowcloud tested against the interface itself
Three statements in this text come from lowcloud's own tests against the interface, not from the maker's documentation. All three differ from it. First, images: over the EU route, Qwen 3.8 Max does not accept images. Alibaba's documentation lists image and video input for its own managed version. Second, fixed JSON schemas: a JSON schema is a field specification the provider enforces so that an answer is guaranteed to contain exactly the fields you defined. Over the EU route the provider rejects such schemas with an error. Alibaba lists structured outputs as supported for its own version. Third, the context window, meaning how much text the model can process in one request. It is measured in tokens; a token is a piece of text roughly half a word to a whole word long. The lowcloud catalogue states 256k tokens, so about 256,000. Alibaba states 1 million for its own managed version. All three differences match the openly published model Qwen3.8-2.4T-A95B: it handles text only, its native context window is 262,144 tokens, and according to its model card its reasoning cannot be disabled. That does not prove this exact model answers over the EU route, it only makes it likely. Plan with what has been measured: no images, no fixed JSON schemas, 256k context.
What the model is good for
Steps where something has to be weighed up. Before answering, the model works through in writing what it needs to take into account. That step is called reasoning and it is always on here. For tasks with several rules, exceptions, or contradictory statements in the source text, this is the difference to a model that starts writing immediately. Concrete places in a workflow: checking a complaint against your own goodwill rules and giving reasons for the decision. Pulling from a long contract the points that deviate from your own template. Putting two quotes side by side and naming where they differ. Reading a vaguely worded customer enquiry and deciding which of the existing workflows is responsible. 256k of context is enough for long single documents and for workflows that carry many intermediate results along. The most sensible use is selective. In lowcloud the model can be set per step: the one step that has to weigh things up runs on Qwen 3.8 Max, the simple steps before and after run on a faster model. That keeps both runtime and cost down.
What it is not good for
No images. If a step has to read a screenshot, a photo, or a scanned invoice, Qwen 3.8 Max is out. Models in the EU catalogue that read images: Kimi K3, Kimi K2.7 Code, Kimi K2.6, GLM 5.3 Flash, Qwen 3.8 27B and Mistral Small 4. No fixed JSON schemas. If a step's answer must be guaranteed to contain certain fields because another system works with it afterwards, use a model that accepts schemas: Kimi K2.6, GLM 5.1, GLM 5, Qwen 3.5 397B or DeepSeek V4 Flash. Qwen 3.5 397B is the closest option from the same series. Not for many small steps. Sorting things into categories, short standard replies, pulling fields out of an email: for these the constant reasoning is too slow, and it produces extra output tokens on every single answer, which you pay for. Not for volumes of text beyond 256k. Listed with 1 million of context in the EU catalogue are Kimi K3, GLM 5.3, GLM 5.3 Flash, GLM 5.2 and DeepSeek V4 Pro.
What it costs
There is no figure here. A markup is added to the purchase price, so a published figure would be wrong. What you get instead: every run shows in the app what it cost in euros. For placement within the EU catalogue: Qwen 3.8 Max sits in the upper middle on price. There are cheaper models and there are more expensive ones. More relevant than that placement is the constant reasoning. It produces extra output tokens on every answer, and output tokens are billed even when the user never sees them. A run with Qwen 3.8 Max therefore ends up costing more in practice than its position in the list suggests. The longer runtime comes on top. The comparison you can rely on runs through your own workflow: execute it once with Qwen 3.8 Max, once with a faster model, and put the two euro amounts side by side. That is also the figure you can show your client.
Try Qwen 3.8 Max
Picking a model needs a free account. You then choose Qwen 3.8 Max in the chat, per agent, or per step.
Common questions
What is Qwen 3.8 Max?
The top tier of the Qwen 3.8 model series, built by Qwen, the AI division of the Chinese company Alibaba. The series was released in August 2026. It is a text model: it reads text and writes text. Over lowcloud's EU route it does not accept images. It works through its reasoning in writing before every answer, and that step cannot be switched off.
Does my data go to China if I use Qwen 3.8 Max?
No. The model was developed in China, but on lowcloud it runs on servers in the EU. The chain: lowcloud UG in Bielefeld passes the request to Melious AI GmbH in Saarbrücken, which runs it in European data centres. Alibaba does not receive the request. Two things are worth saying openly to a client: who built a model is a different question from where it runs. And which European operator serves this particular model is not published by Melious, so lowcloud does not know it either.
Is Qwen 3.8 Max open source?
No. The weights of Qwen 3.8 Max are not published; the model is only usable through an interface. A related model named Qwen3.8-2.4T-A95B is public, and even that is not under MIT or Apache 2.0 but under Qwen's own licence. It requires large products to show the model name visibly, and it requires separate permission for reselling model access above a revenue threshold. Overview pages that write "open source" here are cutting corners.
Can Qwen 3.8 Max read images?
Not through lowcloud. The attempt against the interface fails; lowcloud tested this itself. Alibaba's documentation lists images and video for its own managed version, but that does not hold over the EU route. If a step has to read an image, take Kimi K3, Kimi K2.7 Code, Kimi K2.6, GLM 5.3 Flash, Qwen 3.8 27B or Mistral Small 4 from the EU catalogue.
Why does Alibaba state 1 million tokens of context and lowcloud states 256k?
Because what answers over the EU route does not behave like Alibaba's own managed version. 256k is the figure in the lowcloud catalogue, and that is the one to plan with. It matches the native context window of the openly published model Qwen3.8-2.4T-A95B, which is 262,144 tokens. Whether that exact model sits behind it is not confirmed. If you need more context: Kimi K3, GLM 5.3, GLM 5.3 Flash, GLM 5.2 and DeepSeek V4 Pro are listed with 1 million.
What does Qwen 3.8 Max cost on lowcloud?
The app shows the exact amount on every run, in euros. There is no figure on this page, because a markup is added to the purchase price and the published number would therefore be wrong. For placement: within the EU catalogue the model sits in the upper middle on price. Because it produces extra output tokens for its reasoning on every answer, a run ends up more expensive than that placement suggests.
Qwen 3.8 Max is a product name of Qwen. There is no business relationship between lowcloud and Qwen; we run the openly available model through a European provider. Last updated: 7 September 2026.