Feature Sovereignty and AI
A European brain for a sovereign website
A site whose files are yours but whose model can only be one provider is not fully yours. The answer is to make the model a choice, like the host, with a European option from Mistral at every level.
The challenge
When Dinis Cruz shared the idea of a Sovereign Agentic Website, the answer that came back was a challenge: if you claim the word sovereign, abstract the AI provider, and let people use other models, European ones such as Mistral included. The article agrees, and follows the argument through.
A sovereign website is one you own and can move. One proprietary part in the middle is enough to lock it in, and the model that builds and changes the site is such a part.
If it can only ever be one provider, the site is sovereign in its files and rented in its brain.
The model is a choice
The design is the one already used for hosts: one standard interface, several providers behind it. Mistral offers an OpenAI-compatible endpoint at https://api.mistral.ai/openai/v1, in public preview; its standard /v1/chat/completions uses Mistral’s own, similar format. Ollama and OpenRouter speak the same OpenAI-style API. (Source: Mistral docs, Chat completion and Ollama docs, OpenAI compatibility and OpenRouter, Quickstart; as read on 20 November 2026.)
The model is a choice the owner makes, like the host.
The agents and tools on a sovereign website call that one interface. The owner chooses what sits behind it, at three levels.
1 On the device
A local model on the owner’s own hardware, so a site can be built and changed with nothing leaving the house. Ministral 3 was announced on 2 December 2025 in 3B, 8B and 14B sizes, for edge and local use, each with image understanding, under Apache 2.0. (Source: Mistral AI, Mistral 3; as read on 20 November 2026.)
An estimate, not measured. the 3B and 8B, compressed to about 2 to 5GB, should run on a 16GB Raspberry Pi 5, the 8B slowly; on a tablet, the 3B is the realistic one. (Source: A European brain for a sovereign website (Dinis Cruz, SGit Newsroom); as read on 20 November 2026.) Mistral’s official compressed (Q4_K_M) files are 2.15 GB for the 3B and 5.20 GB for the 8B, plus about 0.85 GB for vision. Those are download sizes, not memory use or speed. (Source: Hugging Face, Ministral-3-3B-Instruct-2512-GGUF and Hugging Face, Ministral-3-8B-Instruct-2512-GGUF; as read on 20 November 2026.)
2 In Europe
Mistral’s API, for the team of agents inside the site. By default, data is hosted in the EEA. (Source: Mistral help centre, Where do you store my data; as read on 20 November 2026.) Hosting is not processing: Mistral does not commit to an inference location on the global endpoint,
api.mistral.ai. Processing in the EU needs the EU endpoint,api.eu.mistral.ai, billed at 1.1 times list price, without Agents, Batch or the Files API, and serving only the models hosted there. (Source: Mistral docs, Regional inference; as read on 20 November 2026.) Mistral’s regional endpoints were announced as generally available on 11 August 2026. (Source: Mistral AI, Regional inference announcement; as read on 20 November 2026.) Mistral’s inference page says “Data sent through our API isn’t used for model training”, but its help centre says that in Free mode it may use API inputs and outputs for training unless you opt out (a toggle in the Admin panel, under Privacy), and that Labs and Public Preview models may be used for training whatever the plan. (Source: Mistral AI, Cloud inference and Mistral help centre, Do you use my data to train and Mistral help centre, Opting out of training; as read on 20 November 2026.)3 Any provider
The owner’s own key, through OpenRouter, for any model, Mistral’s again among them. OpenRouter’s OpenAI-style API is at
https://openrouter.ai/api/v1. (Source: OpenRouter, Quickstart; as read on 20 November 2026.)
What it costs, and what it can do
Ministral 3 8B is listed at 0.15 dollars per million tokens in and 0.15 out. (Source: Mistral docs, Ministral 3 8B and Mistral docs, Pricing; as read on 20 November 2026.) Mistral Small 4 is an Apache 2.0 model with 119 billion parameters, 6.5 billion active, listed at 0.15 dollars per million input tokens and 0.60 per million output tokens. (Source: Mistral docs, Mistral Small 4 and Mistral docs, Pricing; as read on 20 November 2026.) Ministral 3 3B, model name ministral-3b-2512, is listed as Apache 2.0, released 2 December 2025, at 0.10 dollars per million tokens in and 0.10 out. (Source: Mistral docs, Ministral 3 3B and Mistral docs, Pricing; as read on 20 November 2026.)
Mistral offers function calling (Ministral 3 and Mistral Small 4 among the models), and connectors that register remote MCP servers as tools, in public preview. (Source: Mistral docs, Function calling and Mistral docs, Connectors; as read on 20 November 2026.)
What its terms say
Retention. API inputs and outputs are kept for 30 rolling days to monitor abuse, unless zero data retention is on; Agents API data is kept until the account ends. Zero data retention is only for pay-as-you-go, only for stateless calls, must be requested and may be refused, and does not cover Labs or Public Preview models. (Source: Mistral AI, Privacy policy and Mistral help centre, Zero data retention and Mistral docs, Zero data retention; as read on 20 November 2026.)
Where data goes. Some features may move data temporarily outside the EU to listed subprocessors; Enterprise customers can switch some of these off. (Source: Mistral help centre, Where do you store my data; as read on 20 November 2026.)
Certifications. Mistral says it “complies with both SOC 2 Type II and ISO 27001/27701 frameworks”. (Source: Mistral help centre, SOC 2 and ISO 27001; as read on 20 November 2026.)
The licence to watch
Not every open-weight model is equally open. Ministral 3, Mistral Small 4 and Voxtral Small are labelled Apache 2.0. Mistral Large 4 is under a modified MIT licence: its announcement says a company offering it as a model-as-a-service needs a separate commercial licence, while Mistral’s help centre describes the condition as a monthly revenue threshold of 20 million dollars. The pages disagree; read the licence file itself before relying on either. (Source: Mistral docs, Ministral 3 8B and Mistral AI, Mistral Large 4 and Mistral help centre, Licences of the open models; as read on 20 November 2026.) If models are hosted for customers’ sites, that line matters.
Apache 2.0 models first.
Already run, without saying so
On 18 October 2026, Mistral’s free mode was one of four providers that passed the Newsroom’s two tests: free to start, and callable from a web page with the reader’s own key. (Source: Free models, your own key (SGit Newsroom); as read on 20 November 2026.)
On 11 November 2026, at Web Summit, the Swarm ran 216 model calls on four open-weight models (DeepSeek, Llama, Mistral and Qwen), through OpenRouter, for 0.068 dollars. Which Mistral model, and its share of the calls, was not recorded. (Source: Pennies against dollars (SGit Newsroom); as read on 20 November 2026.)
And while this site was being built, on 20 November 2026, its demo agent called ministral-3b-2512 from a terminal and from a browser page; every call and its tokens are in the console.
Working with Mistral: a proposal
Nothing has been agreed with Mistral. The article proposes a conversation: a Mistral option on every site, in the team inside and in the build page, where the same brief can run on Mistral beside other models as an open eval; Ministral 3 on the Pro appliances, measured on the Pi and published; the European stretch goals of the Kickstarter; and credits for the evals and the demo sites.
Mistral lists strategic, ecosystem and integration partners and has a public “Partner with Mistral” expression-of-interest form; its ambassador programme takes rolling applications and offers free API credits, for at least six months’ commitment. (Source: Mistral AI, Partners and Mistral docs, Ambassadors; as read on 20 November 2026.)
Nothing has been agreed with Mistral; this is a proposal.
Not one company either
A European default is still one company. That is why the design is a choice behind one interface, not a new dependency.
Sovereignty cannot rest on one company, European or not.
Mistral as the European default, never the only door.
What this does not claim
- Nothing has been agreed with Mistral. The partnership is a proposal.
- The fit on a Pi or a tablet is an estimate, to be measured. (Dinis Cruz, A European brain for a sovereign website.)
- Terms and prices are as Mistral’s pages showed on 20 November 2026. Where the article and Mistral’s own pages differ (EU processing, training in Free mode, zero data retention, the Large 4 licence, the partner form), this site follows Mistral’s pages. See every source.
Try it yourself
Run the same brief on Mistral beside another model, with your own key or a model on your own machine.