Qwen 3.8 27B
Qwen 3.8 27B is a language model from the Qwen team at Alibaba Cloud in China. It was published on 5 August 2026 under the Apache 2.0 licence, and the weights are openly available on Hugging Face under the id Qwen/Qwen3.8-27B. The model has around 27.8 billion parameters, reads images as well as text, and takes 262,144 tokens per request. Tokens are the pieces a model splits text into. The amount it can take at once is called the context window.
At lowcloud, Qwen 3.8 27B is one of the 14 models that run through Melious AI GmbH on servers in the EU. Your request does not go to Alibaba. You can pick the model per message in chat, fix it for one agent, or set it for a single step of a workflow. Model choice comes with the free account. Without an account you can have a workflow built for you on agent.lowcloud.de, and lowcloud picks the model in that case.
- Vendor
- Qwen, China
- Context window
- 262k Token
- Reads images
- yes
- Licence
- Apache 2.0
- Processing
- EU servers via Melious
Where the model comes from
The maker is the Qwen team at Alibaba Cloud, China. The 27B variant of the 3.8 series has been on Hugging Face since 5 August 2026, with an FP8 build added on 13 August 2026. The licence is Apache 2.0, the standard text unchanged, copyright 2026 Alibaba Cloud. There is no addendum with usage restrictions. Commercial use, modification and redistribution are permitted. That is not true across the whole Qwen range. The two larger models in the same series, Qwen3.8-2.4T-A95B and Qwen3.8-Flash-Next, carry the vendor's own licence and not Apache 2.0. When an overview page states that Qwen is open source across the board, that claim is wrong for those two. Technically, Qwen 3.8 27B is a dense model: all 27.8 billion parameters take part in every answer. Mixture-of-experts models, or MoE, activate only part of their parameters per request. The model has 64 layers and a built-in vision part, so it was designed for text and images from the start rather than extended later.
What "runs on EU servers" actually means
The chain is short and can be named in full: you send the request in lowcloud. lowcloud UG in Bielefeld passes it to Melious AI GmbH in Saarbrücken. Melious runs the model in data centres in the EU. Melious does not operate its own data centres. It distributes load across European operators. Which operator serves Qwen 3.8 27B specifically is not published by Melious. That is the part lowcloud does not know and therefore cannot tell you. If your client needs that detail, say so plainly instead of guessing a name. What is certain: the request does not go to Alibaba. The model is openly available software that a European operator runs on its own hardware. There is no connection to a vendor API in the process.
What the model is good for
Images. The model takes images directly. lowcloud tested this against its own endpoint and it works. That makes steps like these possible: read a photographed or scanned invoice and pull out supplier, amount and date, describe a screenshot of a form, classify a product photo. According to the vendor's model card (August 2026) the model also understands video. lowcloud has not tested that. Long inputs. 262,144 tokens per request cover long contracts, whole email threads, or several files in one step. The model card additionally mentions an extension to up to 1 million tokens. The lowcloud catalogue lists the 262k the model handles natively. Steps that run often. Qwen 3.8 27B is among the cheaper models in the EU catalogue. That makes it a reasonable pick for recurring work: classifying, sorting, summarising, picking the right category out of a piece of text.
What it is not good for
Fixed JSON schemas. Some steps require an answer in a fixed data structure, meaning a JSON schema that states exactly which fields come back with which type. Qwen 3.8 27B does not accept this. The provider rejects such requests with an error. This comes from lowcloud's own test against the API. The vendor's model card says nothing about it. For those steps, use a model from the EU catalogue that does accept fixed schemas: Qwen 3.5 397B from the same maker, Kimi K2.6, GLM 5.1, GLM 5 or DeepSeek V4 Flash. If a step has to read images and return a fixed schema at the same time, Kimi K2.6 is the only option in the EU catalogue. Very short, simple questions. The model reasons before every answer and writes that reasoning as output tokens. Even a one-word answer therefore costs more than with a model that has no reasoning step. The model card states that the reasoning mode can be switched off per request. At lowcloud it stays on. For a step that only has to say yes or no a hundred times a day, this is the wrong model.
What it costs
There are no prices on this page, deliberately. A margin is added on top of the purchase price, so any figure published here would be wrong. The placement: Qwen 3.8 27B is among the cheaper models in the lowcloud EU catalogue. That placement moves up once many short requests add up. The reasoning mode produces extra output tokens, which eat back part of the low token price. With long inputs and little output it barely matters. With short requests in large numbers it does. What a run actually cost is shown in euros on every run in the app. That is the figure to base a price for your client on.
Is the model still current?
Yes, as of 7 September 2026. Qwen 3.8 27B is the newest open model of this size from Qwen. Its predecessor is Qwen3.6-27B from 21 April 2026. There is no successor in the 27B line yet. The 3.8 series has two larger models, both from August 2026: Qwen3.8-2.4T-A95B and Qwen3.8-Flash-Next. Neither is under Apache 2.0, and neither is part of the lowcloud EU catalogue. The EU catalogue does include Qwen 3.8 Max from the same maker. Max has a smaller context window (256k), does not read images, and also does not accept fixed JSON schemas. If you need images, Qwen 3.8 27B is the right one of the two Qwen models in the catalogue.
Try Qwen 3.8 27B
Picking a model needs a free account. You then choose Qwen 3.8 27B in the chat, per agent, or per step.
Common questions
What is Qwen 3.8 27B?
A language model from the Qwen team at Alibaba Cloud in China, published on 5 August 2026 under Apache 2.0. Around 27.8 billion parameters, a 262,144-token context window, reads text and images. At lowcloud it runs on servers in the EU.
Does my data go to China?
No, not when you use the model through lowcloud. The maker is based in China, but the model runs as openly available software in data centres in the EU. The chain is: lowcloud UG in Bielefeld, Melious AI GmbH in Saarbrücken, data centres in the EU. There is no connection to an Alibaba API in the process. Two limits worth telling your client: Melious does not publish which European operator serves the model. And if the same model is used elsewhere through the vendor's own API, this statement does not apply there.
Is Qwen 3.8 27B open source?
The weights are under Apache 2.0, the standard text unchanged and without added conditions. Commercial use is allowed. Training data and training code are not published, so strictly speaking this is an open-weights model. That applies to this model only. The larger models in the same series, Qwen3.8-2.4T-A95B and Qwen3.8-Flash-Next, carry the vendor's own licence.
Can Qwen 3.8 27B read images?
Yes. The vision part is built into the model rather than attached afterwards. lowcloud tested it against its own endpoint: images are accepted and processed. That makes the model suitable for steps involving scans, photos or screenshots.
Why does my workflow fail with an error when I set a fixed JSON schema?
Because Qwen 3.8 27B does not accept fixed JSON schemas. The provider rejects such requests with an error. lowcloud verified this against the API itself; the vendor's model card does not mention it. Use a different model from the EU catalogue for that step: Qwen 3.5 397B, Kimi K2.6, GLM 5.1, GLM 5 or DeepSeek V4 Flash. If the step needs images and a fixed schema at once, Kimi K2.6 is the only option in the EU catalogue.
What does Qwen 3.8 27B cost at lowcloud?
Prices are not listed on this page, because a margin is added on top of the purchase price and any published figure would be wrong. Within the lowcloud EU catalogue the model is among the cheaper ones. The reasoning mode adds output tokens, which pushes costs up when there are many short requests. What a run cost is shown in euros on every run in the app.
Qwen 3.8 27B 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.