Mistral AI: the European AI company

Mistral AI developed unusually fast:

• 2023 – Founded in Paris; Mistral 7B is released in September under the Apache 2.0 licence, followed in December by the mixture-of-experts model Mixtral 8x7B. Both set new standards for open models.
• 2024 – Mistral Large launches as the commercial flagship together with the chat interface Le Chat; Codestral (code), Pixtral (image understanding) and the compact Ministral models for devices and edge deployment follow.
• 2025 – Mistral OCR for documents, the reasoning model Magistral and the Voxtral audio family appear; with Mistral Compute the company also builds its own European AI infrastructure.
• 2026 – La Plateforme offers regional EU endpoints that guarantee processing within Europe.

For dAi Pro this means the default provider now covers every task of the platform – text, embeddings, documents, images and speech. The feature overview shows which building blocks dAi Pro actually uses.

Mistral can be used in three different ways – and the difference matters for organisations:

• Le Chat is the end-user interface, comparable to ChatGPT: an assistant for chatting, research and working with documents – as a web application and app.
• La Plateforme is the API for developers and applications. Systems such as dAi Pro access the models through it: billed per call, with a data processing agreement, tenant controls and EU endpoints.
• Self-hosting: the open Mistral models can run on your own hardware or with European GPU providers – the most sovereign but also the most demanding option. Mistral models are also available in the major cloud catalogues (such as Azure).

dAi Pro takes the second route: the La Plateforme API. Your knowledge base stays in your own PostgreSQL database – only the question and the relevant text passages are sent to Mistral, never the entire corpus.

A few technical properties explain why Mistral models work so well in practice:

• Mixture of Experts (MoE) – models like Mixtral activate only part of their parameters ("experts") per request. That reduces compute and response time while keeping quality high – one reason for the short latencies our RAG comparison measures continuously.
• Large context window – Mistral Large processes up to 128,000 tokens of context. dAi Pro's CAG strategy benefits in particular, as it passes entire knowledge bundles along directly.
• Function calling – the models can invoke tools in a structured way. dAi Pro uses this for its HTTP and SQL tools, for example to look up prices or stock levels live.
• Structured outputs – schema-conform JSON answers make the models reliable for classification and data extraction; dAi Pro uses this in the query classifier that decides between RAG and CAG, among other places.
• Fine-tuning – La Plateforme offers a fine-tuning API; in practice dAi Pro usually achieves adjustments more economically via assistant profiles and example answers.

What does GDPR-compliant operation look like in concrete terms when an organisation runs dAi Pro with Mistral?

• Data processing agreement – operator and Mistral sign a DPA under Art. 28 GDPR with a French contracting party; standard contractual clauses for third-country transfers are not required.
• Data minimisation by design – per request, dAi Pro transmits only the question and the matching text passages. Original documents, vectors and logs stay in your own infrastructure (PostgreSQL with pgvector).
• Retention – Mistral retains API data only briefly by default for abuse monitoring (30 days); zero-retention agreements are available for enterprise customers. On the operator side, dAi Pro deletes chat logs automatically after 90 days.
• Transparency – the chat identifies itself to visitors as an AI (EU AI Act) and discloses the provider in its help panel.

The overall architecture is described on GDPR-compliant AI.

dAi Pro uses four Mistral building blocks, each in a clearly defined place:

• Chat (Mistral Large) – answers visitors' questions. dAi Pro's hybrid search supplies the matching passages, and strict grounding ensures answers come only from the knowledge base. Fallback models step in if the service is disrupted.
• Embeddings (mistral-embed) – translates every knowledge chunk into a 1,024-dimensional vector. These vectors live in your pgvector database and enable semantic search: what is meant is found – not just what is written verbatim.
• Documents (Mistral OCR) – reads PDFs and scans as structured Markdown, including tables. dAi Pro also extracts figures as separate image documents and describes them with AI – making diagrams searchable.
• Speech (Voxtral) – recognises spoken questions in real time and reads answers aloud, streamed as they are generated. Technical terms can be corrected for pronunciation.

All four building blocks are configurable per collection – if you wish, you can combine Mistral embeddings with a different chat model. The feature list shows how they interact.

La Plateforme bills per processed token – input and output separately, with much lower rates for smaller models and embeddings. Compared with the US flagships, Mistral Large is consistently among the less expensive top models; our RAG comparison puts the current token prices of all four providers side by side.

More important than the list price is control: dAi Pro logs every single call with its token usage and converts the costs into euros – analysable per collection and period, exportable for accounting. On top of that, the answer cache and the CAG strategy noticeably reduce the number of expensive calls: recurring questions cost nothing at all, and frequent topics answer from pre-warmed knowledge.