DeepSeek V4 Flash
DeepSeek V4 Flash is a language model built by DeepSeek, a company in China. The current revision is called DeepSeek-V4-Flash-0731; the four digits are the date, 31 July 2026. The weights, meaning the trained model file, have been public on Hugging Face since 1 August 2026 under the MIT licence. It is an MoE model with roughly 304 billion parameters.
At lowcloud, DeepSeek V4 Flash runs through Melious AI GmbH in Saarbrücken on servers inside the EU. The request never goes to DeepSeek in China. One operating fact matters more than the rest: V4 Flash is the default. A workspace where nobody picked a model runs on this one. So if you build something at lowcloud and change nothing, you are using DeepSeek V4 Flash.
- Vendor
- DeepSeek, China
- Context window
- —
- Reads images
- no
- Licence
- MIT
- Processing
- EU servers via Melious
Maker, licence, revision
DeepSeek is an AI company in China. The V4 Flash line has been public since June 2026, the base model since April 2026. The 0731 revision described here has been on Hugging Face since 1 August 2026. The licence is MIT. That is a plain open source licence with no conditions on use: anyone may download the weights, run them, change them and use them commercially. This is not a given for Chinese models. Some are released under a custom licence with restrictions while overview sites still label them "open source". For DeepSeek V4 Flash the label is correct. Technically V4 Flash is an MoE model. MoE stands for Mixture of Experts: the model is made of many sub-networks, and only a small share of them does the work for each token. A token is a piece of text, roughly half a syllable to a whole one. The configuration file lists 256 such experts, six of which are active per token. DeepSeek does not publish how many parameters that adds up to per token. There is no successor to this revision. In early September 2026 DeepSeek added a variant that accepts images (DeepSeek-V4-Flash-Vision-Exp). It is marked experimental and is not in the lowcloud EU catalogue. The larger sibling DeepSeek V4 Pro is a separate size, not a successor.
The default model at lowcloud
If you pick no model, you get DeepSeek V4 Flash. That applies to every workspace where the model setting was left alone. Internally the setting is called EU_AUTO_MODEL. You can change it in three places: per message in the chat, fixed for one agent, or for a single step of a workflow. An agent can stay on V4 Flash for the simple steps and switch to another model only for the one step that needs it. On agent.lowcloud.de you can have a workflow built without an account. In that case lowcloud picks the model. Choosing the model yourself comes with the free account.
What "on EU servers" actually means
The chain is: customer, then lowcloud UG in Bielefeld, then Melious AI GmbH in Saarbrücken, then data centres inside the EU. DeepSeek the company is not part of that chain and sees neither the request nor the answer. This works because the weights are public, so any operator can run them on their own hardware. Melious runs no data centres of its own. It distributes requests across European operators. Which operator serves a given model is not published by Melious. That also states what lowcloud does not know: the specific site and operator behind a single request. If a client asks, that is the honest answer. What can be committed to is processing inside the EU, not the individual operator.
What the model is good for
V4 Flash is the cheapest tier in the EU catalogue. That makes it a sensible default for workflows that run often or consist of many steps: sorting incoming items, summarising text, pulling fields out of documents, drafting replies, handing results to the next step. The model accepts fixed JSON schemas. JSON is a data format with named fields. A fixed schema means you state which fields the answer must contain and the model sticks to it. In a workflow that is the deciding point, because otherwise the next step does not know where the invoice number, the date and the amount are. lowcloud tested this against the API itself rather than copying it from the maker's page. DeepSeek's own guides document JSON output only on the larger V4 Pro model. The lowcloud catalogue lists no context window for this model. The context window is how much text a model can read at once in a single run. The model's public configuration file states 1,048,576 tokens, about one million. Whether the EU servers offer that full length has not been verified.
What it is not good for
The model cannot read images. Invoices as photos, screenshots, scans or ID documents are out of reach. For those steps switch to a model in the EU catalogue that reads images: Kimi K3, Kimi K2.7 Code, Kimi K2.6, GLM 5.3 Flash, Qwen 3.8 27B or Mistral Small 4. If very long documents have to be processed in one piece, a model with a stated length is the safer pick, for example Kimi K3, GLM 5.3 or DeepSeek V4 Pro, each listed at one million tokens. The cheapest tier is not automatically right for the hardest decision in a workflow. If one step keeps getting it wrong, move that step to a larger model and leave the rest on V4 Flash. This page quotes no performance figures. Benchmark numbers published by a maker say little about how a model decides inside your specific workflow. The only reliable check is running the same workflow on two models and comparing the output.
What it costs
No prices here. A markup is added to the purchase price, so any figure published on this page would be wrong. The relative position can be stated: DeepSeek V4 Flash is the cheapest of the 14 EU models at lowcloud. Every other model in the EU catalogue costs more. The actual cost in euros is shown on every single run in the app. So before you roll a workflow out to a client you can see what one pass costs, and you can move individual steps to a more expensive model without raising the cost of the whole workflow.
Try DeepSeek V4 Flash
Picking a model needs a free account. You then choose DeepSeek V4 Flash in the chat, per agent, or per step.
Common questions
Does my data go to China?
No. The request does not go to DeepSeek. It goes through lowcloud UG in Bielefeld to Melious AI GmbH in Saarbrücken and from there into data centres inside the EU. The model is a publicly released file that European operators run on their own servers. The maker in China sees neither the request nor the answer. That covers the model call. If your agent also calls other services, those have their own routes.
Which model does lowcloud use if I pick none?
DeepSeek V4 Flash, revision 0731. It is the default for every workspace with no model choice of its own. You can change it per message in the chat, fixed per agent, or per single step of a workflow.
Is DeepSeek V4 Flash open source?
Yes. The weights are published on Hugging Face under the MIT licence and may be used commercially with no conditions attached. That is not always the case with Chinese models; some carry a custom licence with restrictions. This one does not.
Can DeepSeek V4 Flash read images?
No. lowcloud verified this against the API itself. DeepSeek released an experimental variant with image input in September 2026, and it is not in the EU catalogue. For images use 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 DeepSeek V4 Flash?
The lowcloud catalogue states no figure. The model's public configuration file lists 1,048,576 tokens, about one million. Whether the EU servers offer that length in full has not been verified. If length is critical, pick a model the catalogue gives a number for.
What does one run with DeepSeek V4 Flash cost?
lowcloud publishes no fixed figures, because a markup is added to the purchase price. Within the EU catalogue V4 Flash is the cheapest tier. The actual cost in euros is shown on every single run in the app.
DeepSeek V4 Flash 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.