Qwen 3.5 397B

Qwen 3.5 397B is a language model from the Chinese company Alibaba, built by its AI team Qwen. It was released in February 2026 under the Apache 2.0 licence, and the weights are openly downloadable on Hugging Face. The number in the name is the size: 397 billion parameters are stored, of which only around 17 billion are used per token. A token is the smallest unit of text a model processes, usually a short word or part of a word.

At lowcloud, Qwen 3.5 397B is one of 14 models that run on servers in the EU. Operation goes through Melious AI GmbH in Saarbrücken, and the request does not go to Alibaba. You can pick the model per message in chat, set it permanently for one agent, or assign it to a single step of a workflow. One point belongs before you use it: Qwen describes the model as able to read images, and the Melious model page lists it that way too. lowcloud tested this against the interface. The model does not reject an image, it returns an empty answer. lowcloud therefore lists it as not image-capable.

Vendor: Qwen, China
Context window: 262k Token
Reads images: no
Licence: Apache 2.0
Processing: EU servers via Melious

Where the model comes from

It was built by Qwen, the AI team of the Chinese company Alibaba, based in Hangzhou. The repository on Hugging Face was created on 16 February 2026. The full release name is Qwen3.5-397B-A17B. The A17B stands for the active parameters: 397 billion are stored in total, around 17 billion of them are used per token. The licence is Apache 2.0, a free licence with no extra conditions. Downloading, running, modifying and commercial use are all permitted, with no cap on user numbers and no approval required. That is not a given: several large models listed as open source on overview sites carry custom licences with conditions attached. This one does not. Qwen does not publish the training data, though, so open weights is the more accurate term than open source. On the build: the model has 60 layers and is a Mixture of Experts, MoE for short. It consists of 512 sub-networks, of which only eleven are used per token: ten that change and one that is always involved. The weight files come to roughly 800 gigabytes.

What "runs on EU servers" means here

The chain can be named in full. The customer talks to the agent. The agent runs at lowcloud UG in Bielefeld. lowcloud passes the request to Melious AI GmbH in Saarbrücken. Melious runs the model in data centres in the EU. Alibaba is not part of this chain. The model is freely licensed, so a European provider can run the weights itself. No text is sent to a server belonging to the model's maker. What lowcloud does not know: Melious operates no data centres of its own, it distributes across European operators. Melious does not publish which operator serves a given model. If a client asks for the specific location, the answer you can stand behind is: inside the EU, the contracting party for operation is Melious in Saarbrücken, the individual operator is not disclosed.

What the model is good for

Long text. The context window is the amount of text a model can take into account at once. In the lowcloud catalogue it is 262,000 tokens. That covers long contracts, sets of minutes or entire email threads in one step, without splitting them up first. Steps with a fixed data format. The model accepts fixed JSON schemas. JSON is a machine-readable data format. A fixed schema means the fields are defined in advance and the model has to fill exactly those fields. That is the difference between a step that writes reliably into a spreadsheet or a CRM and one you clean up afterwards. lowcloud tested this against the interface rather than copying it from the maker's page. Multilingual workflows. Qwen's model card states support for 201 languages and dialects. That is the maker's claim, not a measurement by lowcloud. Tool calls. Qwen describes the model as strong at calling tools, meaning it triggers steps such as searching, calculating or querying data. That is also the maker's claim. One practical note on how it runs: by default the model first produces an intermediate text with its reasoning and only then the actual answer. That costs extra tokens and time.

What it is not good for

Images. This is the most important point. The Hugging Face model card lists an image encoder, and the Melious model page lists the model as image-capable. On the EU interface something else happens: send an image and you get no error message, you get an empty answer. That is worse than an error, because the workflow step formally completes and still delivers nothing. lowcloud tested this and therefore lists the model as not image-capable. This observation is lowcloud's own and contradicts both of the published statements. If a workflow has to read images, the EU catalogue holds six alternatives: Kimi K3, Kimi K2.7 Code, Kimi K2.6, GLM 5.3 Flash, Qwen 3.8 27B and Mistral Small 4. If you need images and fixed JSON schemas together, exactly one of them fits: Kimi K2.6. If you want to stay with the same maker, take Qwen 3.8 27B. Input beyond 262,000 tokens. The catalogue has models with a one million token window for that: Kimi K3, GLM 5.3, GLM 5.2 and DeepSeek V4 Pro. Short steps that run very often. The reasoning text and the size of the model make it the more expensive choice here compared with a small model.

The model has been superseded

Qwen has released two newer generations since. Qwen3.6 arrived in April 2026, Qwen3.8 in August 2026. From the newer generation, the lowcloud catalogue holds Qwen 3.8 Max and Qwen 3.8 27B. Qwen 3.8 27B reads images and has the same 262,000 token context window. Qwen 3.8 Max has 256,000 tokens and, according to lowcloud's testing, does not read images. Superseded does not mean switched off. Qwen 3.5 397B stays in the catalogue and keeps working. An agent that runs reliably on it does not need to be moved. For new agents it is worth checking the 3.8 models first, because that is where Qwen continues development. Neither of them is a replacement in the same size class: Qwen 3.8 27B is considerably smaller.

What it costs

There are no prices here. The reason: a margin is added on top of the purchase price, so any figure published here would be wrong. What can be said: within lowcloud's EU catalogue, Qwen 3.5 397B sits in the lower middle. It is neither the cheapest nor the most expensive model. A model this large not sitting at the top is down to the MoE build: only around 17 of the 397 billion parameters are used per token. The actual cost is shown on every run. Each run of an agent displays its amount in euros in the app. That lets you work out what one case costs before rolling the workflow out to a client, instead of estimating it. One cost factor is the intermediate text with the model's reasoning. It consists of tokens as well and therefore shows up in the amount for the run.

Try Qwen 3.5 397B

Picking a model needs a free account. You then choose Qwen 3.5 397B in the chat, per agent, or per step.

Book a call

Common questions

Does my data go to China if I use Qwen 3.5 397B?

No. The model comes from China, the processing does not. The weights are freely licensed and are run by Melious AI GmbH in Saarbrücken in data centres in the EU. The chain is: customer, lowcloud UG in Bielefeld, Melious AI GmbH in Saarbrücken, data centres in the EU. Alibaba receives no request. What is not known is which European operator serves this specific model, as Melious does not publish that. If you use Qwen directly through Alibaba Cloud or qwen.ai, this chain does not apply.

Is Qwen 3.5 397B open source?

The licence is Apache 2.0, a free licence with no extra conditions. Downloading, running, modifying and commercial use are permitted, with no user cap and no approval. Qwen does not publish the training data, so open weights is a more accurate term than open source.

Can Qwen 3.5 397B read images?

Not at lowcloud. Qwen's model card lists an image encoder, and the Melious model page lists the model as image-capable. But when tested against the EU interface, the model answers an image with an empty response instead of rejecting it. lowcloud therefore lists it as not image-capable. For images, the catalogue offers Kimi K3, Kimi K2.7 Code, Kimi K2.6, GLM 5.3 Flash, Qwen 3.8 27B or Mistral Small 4.

How large is the context window of Qwen 3.5 397B?

262,000 tokens according to the lowcloud catalogue. That matches Qwen's stated native context of 262,144 tokens. The model card additionally mentions an extension to around one million tokens using scaling techniques. That extension is not what runs at lowcloud; the catalogue figure applies here.

What does 397B in the name stand for?

397 billion parameters. The full release name is Qwen3.5-397B-A17B, where A17B stands for 17 billion active parameters per token. The model is built as a Mixture of Experts, MoE for short: it is split into 512 sub-networks, of which only eleven are used per token.

Is there a successor to Qwen 3.5 397B?

Yes. Qwen released Qwen3.6 in April 2026 and Qwen3.8 in August 2026. The lowcloud catalogue holds Qwen 3.8 Max and Qwen 3.8 27B from those. Qwen 3.5 397B remains usable and keeps running. If you are building a new agent, check the 3.8 models first.

Qwen 3.5 397B 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.