Mistral Small 4: built in France, processed in the EU
Mistral Small 4 is a language model from Mistral AI in Paris. The company released it on 16 March 2026 under the Apache 2.0 licence. The model has 119 billion parameters but only computes with around 6 billion of them per token. A token is the text fragment a sentence is split into before processing. The model reads text and images and has a context window of 256k tokens, meaning the amount of text it can take in during one pass. The API name is mistral-small-2603. The weights, the trained numbers that make up the model, are published openly on Hugging Face.
At lowcloud, Mistral Small 4 is one of 14 models that run through Melious AI GmbH in Saarbrücken on servers inside the EU. The request does not go to Mistral. In the lowcloud EU catalogue, Mistral Small 4 is the only model whose maker is based in Europe. The other thirteen come from makers in China. If you have to explain to a client not just where the server stands but where the model comes from, this is the one option in the catalogue.
- Vendor
- Mistral AI, Frankreich
- Context window
- 262k Token
- Reads images
- yes
- Licence
- Apache 2.0
- Processing
- EU servers via Melious
Where the model comes from
The maker is Mistral AI, based in Paris, France. Mistral Small 4 was released on 16 March 2026. The digits in the API name mistral-small-2603 stand for that date: year 26, month 03. The licence is Apache 2.0. This is a genuine open source licence with no conditions on company size, revenue or purpose of use. Commercial use, redistribution and modification are permitted. That is not a given with language models: many overview pages write "open source" even when the licence carries conditions. For Mistral Small 4 it is the standard Apache 2.0, as stated in the model card on Hugging Face and in Mistral's own documentation. The weights are available there for download under the name mistralai/Mistral-Small-4-119B-2603. The architecture is a mixture of experts, or MoE. The model consists of 128 sub-networks. For each token, only 4 of them are active. That is why the model has 119 billion parameters in total but only has to compute around 6 billion of them per token. Fewer active parameters means less computation per token. Mistral lists 23 supported languages, among them German, French, English, Spanish and Italian. No benchmark figures appear here. The numbers in the model card come from the maker itself; we have no independent measurements.
What "on EU servers" means in practice
The chain is: your client sends the request to the agent. The agent runs at lowcloud UG in Bielefeld. lowcloud passes the request to Melious AI GmbH in Saarbrücken. Melious operates no data centres of its own and distributes requests across European operators. The model runs there. The request does not go to Mistral. This works because the weights are openly available under Apache 2.0: any operator is allowed to run the model itself. So you are talking to a model built in France without data reaching its maker. What lowcloud does not know: which specific data centre operator serves an individual request. Melious does not publish that assignment. If a client needs the exact location written into a contract, that is the point where you have to ask. What can be stated: processing inside the EU, contracting party in Germany, no path to the model's maker.
What it is good for in a workflow
Mistral Small 4 is a sensible default for steps that run often and carry no special requirement: summarising text, classifying it, rewriting it, translating it, pulling details out of an email or a document, drafting a reply. It reads images. That makes steps possible which begin with a photo or a scan: a photographed invoice, a screenshot from a ticket, a scanned form. Six of the 14 models in the EU catalogue can do this: Mistral Small 4, Kimi K3, Kimi K2.7 Code, Kimi K2.6, GLM 5.3 Flash and Qwen 3.8 27B. The context window of 262k tokens in the lowcloud catalogue is enough for long documents in a single pass, roughly several hundred pages of text. And it is the answer when a client asks about European origin. Not only does the server stand in the EU, the maker is based here as well.
What it is not good for
Mistral Small 4 does not accept fixed JSON schemas over the EU route. A JSON schema is a template that defines exactly which fields the answer must contain. lowcloud tested this against the interface: the request is rejected with an error. That is lowcloud's own observation, not a statement from the maker; Mistral's documentation describes structured output as generally possible. If a step in your workflow needs a fixed field format because another system processes the answer further, use one of the models that accept schemas: Kimi K2.6, GLM 5.1, GLM 5, Qwen 3.5 397B or DeepSeek V4 Flash. You can still ask for JSON in the prompt text, but the format is then not enforced. For texts above 262k tokens the context window is not enough. Kimi K3, GLM 5.3, GLM 5.3 Flash, GLM 5.2 and DeepSeek V4 Pro work with one million tokens. Mistral Small 4 sits in the cheapest price tier of the catalogue. For steps where the quality of the answer decides the outcome, run the same workflow against a larger model and compare before you commit to it in front of a client.
What it costs
No amounts appear on this page. Any figure printed here would not be the price you end up paying. Placement in the EU catalogue: Mistral Small 4 sits in the cheapest tier, together with the flash models GLM 5.3 Flash and DeepSeek V4 Flash. The actual cost is shown in euros on every single run inside the app. That lets you run a workflow once with Mistral Small 4 and once with another model and compare the two amounts directly before you quote a price to your client.
Has Mistral Small 4 been superseded?
No, not as of 7 September 2026. Mistral's news page lists no newer model in the Small line. The predecessor is Mistral Small 3.2 from June 2025. Mistral deprecated the older versions Mistral Small 3.1, 3.0 and 2.0 on 30 November 2025. According to the maker, Mistral Small 4 combines the capabilities of several earlier models in one, including the chat style of Mistral Small 3.2. On 22 May 2026 Mistral released Mistral Medium 3.5. That is a different, larger model line and is not part of the lowcloud EU catalogue. Only the maker's own sources were checked for this. A broad web search was not available during research.
Try Mistral Small 4
Picking a model needs a free account. You then choose Mistral Small 4 in the chat, per agent, or per step.
Common questions
Who built Mistral Small 4 and when was it released?
Mistral AI, based in Paris, France. Released on 16 March 2026. The API name mistral-small-2603 carries that date.
Is Mistral Small 4 really open source?
Yes. The licence is Apache 2.0, a genuine open source licence with no conditions on revenue, user count or purpose. Commercial use is permitted. The weights can be downloaded from Hugging Face under mistralai/Mistral-Small-4-119B-2603.
Does my data go to China, or to Mistral in France?
Neither. Processing runs from lowcloud UG in Bielefeld to Melious AI GmbH in Saarbrücken and from there onto data centres inside the EU. The request never reaches the model's maker. At lowcloud that also holds for the models built by Chinese makers in the catalogue; with Mistral Small 4 the maker itself is additionally based in the EU. Which operator provides a given data centre is not published by Melious.
Can Mistral Small 4 read images?
Yes. It accepts text and images and answers with text, so workflows can start from a photo, screenshot or scan. lowcloud verified this against the interface.
Why are fixed JSON schemas rejected with Mistral Small 4?
The provider rejects requests carrying a fixed JSON schema over the EU route with an error. lowcloud tested this directly; the maker's documentation does not contradict it but describes structured output as generally possible. If your workflow needs a fixed field format, use Kimi K2.6, GLM 5.1, GLM 5, Qwen 3.5 397B or DeepSeek V4 Flash.
How large is the context window of Mistral Small 4?
The lowcloud catalogue states 262k tokens, Mistral states 256k tokens. Both describe the same order of magnitude, roughly several hundred pages of text in one pass. For longer texts, pick a model with one million tokens: Kimi K3, GLM 5.3, GLM 5.3 Flash, GLM 5.2 or DeepSeek V4 Pro.
Mistral Small 4 is a product name of Mistral AI. There is no business relationship between lowcloud and Mistral AI; we run the openly available model through a European provider. Last updated: 7 September 2026.