Model catalogue · Mistral AI
Mistral
Version via EU inference of open models through the Gosign chain: Mistral 4 119B , model mistralai/Mistral-Small-4-119B-2603 (as of 29 September 2026)
EUR 0.30 per million tokens as of 12 October 2026
Vendor: mistral-small-2603, as of 29 September 2026
Mistral Small 4 from the French vendor Mistral: a mixture-of-experts model with 119 billion parameters, of which 6 billion work per token (8 billion including the embedding and output layers), with a long context, text and image input and open weights under the Apache licence. It carries drafting, analysis and correspondence work as well as work in series, such as pulling mandatory details from incoming documents, works cleanly in German and French and is the obvious choice when the model vendor is to come from Europe as well. From day one it runs through the Gosign chain, as one of the open models Gosign hosts itself in Germany: there the addendum under section 203 of the German Criminal Code is signed, the confidentiality undertaking exists in text form with instruction on criminal liability and continues as a chain obligation onto the subcontractors; inference stays inside the EU, with no US group in the chain. Some of the other routes of this entry still run the older tier Small 3.
Everything section 203 requires is settled, including the provider's confidentiality undertaking, and nothing is stored at the provider, apart from a technical cache of a few minutes for Claude (zero retention). You do not have to clarify anything further.
The operator bills separately - input EUR 0.20 · output EUR 0.60. The figure above is the blend of the two, weighted three quarters input to one quarter output, as it arises in a firm. Charged at the vendor's list price, without a mark-up.
The providers set these prices; they can change at any time, and our figures may contain errors. The list price in force at the time is authoritative.
What it is good for
- Mixture-of-experts with 119 billion parameters, 6 billion active per token - fast enough for work in series, large enough for drafting and analysis
- European model vendor with open weights under Apache 2.0 - a European chain end to end
- Long context of around 256,000 tokens and image input, for example for scanned attachments; open weights under Apache 2.0
- EU inference with the operator's section 203 addendum from day one
Where it stops
- Case-law citations and references must be verified as a matter of principle - invented citations are a documented failure mode of generative models
- No legal or tax-specific fine-tuning - the professional assessment remains entirely with the practitioner
- For the hardest drafting and analysis tasks the flagship Mistral Large is the stronger choice
- Not every route of this entry delivers Small 4 - which tier an instance gets is stated with the version of the respective route
- The platform's global endpoint gives no residency guarantee - the same models can be called there as well. Binding to the EU is therefore a configuration requirement: the global endpoint is blocked by organisational policy, otherwise the EU commitment is a settings question rather than a property.
Status through us
The same model, procured differently, is a different legal position. The status above refers to procurement through us; processing takes place in Germany.
Which route carries it, with which instrument and under which condition, we set out in full to the firm before it commissions us - section 203(4) requires that the chain can be checked, not that it is published. The full matrix, with the sources behind every route, can be requested in the German catalogue after confirming your e-mail address, and we go through it in the conversation.
Is Mistral the right fit?
We go through your tasks and the route that carries them - and say plainly where it does not.
Arrange a conversation