DeepSeek V4 Pro: what the model does and where it runs
DeepSeek V4 Pro is a language model from the Chinese company DeepSeek. The preview version was released on 26 April 2026, the official version carrying the revision number 0813 in mid-August 2026. The preview has 1.6 trillion parameters, of which around 49 billion do the work on any single request; revision 0813 comes to 1.7 trillion. The context window holds 1 million tokens. Context window means: that is how much text the model can read at once in one request. A token is a piece of text of roughly four characters. The weights, meaning the model file itself, are published under the MIT licence and can be downloaded by anyone.
At lowcloud, DeepSeek V4 Pro is one of 14 models that run on servers in the EU. You pick it in the chat for a single message, fix it for one agent, or set it for one single step of a workflow. The request does not go to DeepSeek in China. It goes through lowcloud UG in Bielefeld and Melious AI GmbH in Saarbrücken to data centres in the EU. Choosing the model is part of the free account. Without an account you can have a workflow built for you at agent.lowcloud.de, and lowcloud picks the model in that case.
- Vendor
- DeepSeek, China
- Context window
- 1 Mio. Token
- Reads images
- no
- Licence
- MIT
- Processing
- EU servers via Melious
Origin, licence and architecture
DeepSeek is a company based in China. The preview version of DeepSeek V4 Pro appeared on 26 April 2026. The official version carries the revision number 0813 and was published in mid-August 2026. It replaces the preview. The architecture is a mixture-of-experts model, MoE for short. That means the model is split into many sub-networks, and only a small part of them runs on any single request. In the preview version that is 49 billion out of 1.6 trillion parameters. Revision 0813 reaches 1.7 trillion parameters because it includes an extra module called DSpark that speeds up output. How many parameters are active per request in that revision is not stated on the model card. The model has three thinking levels: low, high and max. They control how long the model deliberates before answering. For the two higher levels the maker recommends a maximum output length of 384,000 tokens. The licence is the MIT licence. In this case the label open source is correct: the weights are public on Hugging Face, and commercial use and redistribution are allowed with no extra conditions. That is not the norm. For other models, overview pages often write open source even though a custom licence with conditions applies. For DeepSeek V4 Pro the label holds. There is no successor. 0813 is the current state of DeepSeek V4 Pro. DeepSeek also has a smaller line called V4 Flash, which is available at lowcloud as well. Since early September 2026 that line has an experimental version that can read images; it is a different model and is not available at lowcloud.
What "on EU servers" means here in concrete terms
The chain is: you send the request to lowcloud. lowcloud UG in Bielefeld passes it to Melious AI GmbH in Saarbrücken. Melious runs it in data centres in the EU. DeepSeek never receives the request. The maker only supplies the weights, meaning the file the model consists of. Whoever runs that file determines where the data goes. There is one point lowcloud does not know either: Melious operates no data centres of its own and distributes work across European operators. Melious does not publish which operator serves this particular model. If your client needs that detail, it has to be requested from Melious. Everything else in the chain you can name to your client: two German companies, and execution inside the EU.
The DeepSeek app, the maker's API and the open weights are three different things
Three things get mixed up with DeepSeek. For a conversation with a client, the difference is the important part. First, the DeepSeek app on the phone. On 27 June 2025 the Berlin Commissioner for Data Protection and Freedom of Information reported it to Apple and Google as unlawful content. The stated reason: unlawful transmission of users' personal data to China. Apple and Google decide themselves whether to block it. Whether the app is currently available in German app stores is something we have not checked. Second, DeepSeek's own programming interface. Anyone calling it directly sends their data to the company in China. Third, the open weights. They are a file under the MIT licence that anyone can download and run on their own servers. That is the route lowcloud takes. The same model does the computing, but on servers in the EU. Neither the app nor the maker's interface is involved. The news coverage of DeepSeek in Germany almost always concerns the app or the maker's interface. It says nothing about where data goes when the weights are run in Europe.
What the model is good for in the builder
The long context is the practical reason to pick this model. 1 million tokens amount to more than a thousand pages of text in a single request, roughly counted. You do not have to cut large documents into pieces first. Tasks it fits: summarising or comparing long documents, finding single details inside a large body of text, multi-step workflows that carry many intermediate results along, and text work with a lot of background material. The maker names improved agentic ability as the focus of revision 0813, meaning tasks where the model calls several tools one after another. That is a maker's claim. We have no measurements of our own on it. The model can be set per step. You can use DeepSeek V4 Pro for the one step that reads the long document and pick a smaller model for the remaining steps.
What it is not good for
DeepSeek V4 Pro cannot read images. If your workflow involves screenshots, photos or scanned pages, you need a different model. In the EU catalogue, Kimi K3, Kimi K2.7 Code, Kimi K2.6, GLM 5.3 Flash, Qwen 3.8 27B and Mistral Small 4 can read images. The model does not accept fixed JSON schemas. A fixed schema is a specification that states exactly which fields the answer must contain. We tested this against the interface: the request is rejected with error 400. What the maker documents is only a general JSON mode that forces valid JSON but no particular shape. This observation comes from our own test, not from the maker's page. If you still need fixed fields, there are two routes: describe the wanted format in the instruction text and check the answer afterwards, or use a model that accepts schemas. In the catalogue those are Kimi K2.6, GLM 5.1, GLM 5, Qwen 3.5 397B and DeepSeek V4 Flash. One last point is not a technical one. If your client rejects a Chinese model on principle, execution in the EU will not settle the matter. In the catalogue only Mistral Small 4 comes from France; every other model comes from a Chinese company.
What it costs
No prices are listed here. A margin is added to the purchase price, so any figure published on this page would be wrong. What you can rely on for orientation: DeepSeek V4 Pro sits in the middle of the 14 EU models on price. It is neither the cheapest nor the most expensive. The actual cost in euros is shown on every single run in the app. You see it after the run and can pass it on to your client. A model of this size with long thinking phases produces more output than a small model, and output drives a large share of the cost. For steps that need no long deliberation, a smaller model is the cheaper choice.
Try DeepSeek V4 Pro
Picking a model needs a free account. You then choose DeepSeek V4 Pro in the chat, per agent, or per step.
Common questions
What is DeepSeek V4 Pro?
A large language model from the Chinese company DeepSeek. The preview was released on 26 April 2026, the official version as revision 0813 in mid-August 2026. The preview has 1.6 trillion parameters with around 49 billion active per request; revision 0813 comes to 1.7 trillion. The context window holds 1 million tokens. The weights are published on Hugging Face under the MIT licence.
Does my data go to China?
No, not when you use the model through lowcloud. Your request goes to lowcloud UG in Bielefeld, from there to Melious AI GmbH in Saarbrücken, and it runs in data centres in the EU. DeepSeek never receives it; the maker only supplies the model file. It is different with the DeepSeek app and with DeepSeek's own programming interface: there the data goes to the company in China. That is why the Berlin data protection commissioner reported the app to Apple and Google as unlawful content on 27 June 2025.
Is DeepSeek V4 Pro open source?
Yes, here the label is accurate. The weights are published on Hugging Face under the MIT licence. Commercial use and redistribution are allowed with no extra conditions. For many other models, overview pages say open source even though a custom licence with conditions applies. For this one it is the MIT licence.
Can DeepSeek V4 Pro read images?
No. The model handles text only. For images, pick 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. Since early September 2026 DeepSeek has an experimental image version of the smaller Flash line, which is not available at lowcloud.
Why does DeepSeek V4 Pro reject my fixed JSON schema?
Because the provider does not support it, even though its documentation talks about JSON output. A fixed schema is rejected with error 400. We tested this ourselves against the interface. What is documented is only a general JSON mode with no schema. If your workflow needs fixed fields, use Kimi K2.6, GLM 5.1, GLM 5, Qwen 3.5 397B or DeepSeek V4 Flash.
What does DeepSeek V4 Pro cost at lowcloud?
We do not name a figure, because a margin is added to the purchase price. For orientation: DeepSeek V4 Pro sits in the middle of the 14 EU models on price. The actual cost in euros is shown on every single run in the app.
DeepSeek V4 Pro is a product name of DeepSeek. There is no business relationship between lowcloud and DeepSeek; we run the openly available model through a European provider. Last updated: 7 September 2026.