API integration for Mistral
We build your Mistral connector
We wire Mistral into your product to Mistral for text, OCR and function calling, with inference on api.eu.mistral.ai when the contract requires it. api.mistral.ai is not Europe.
- Senior product team
- LLM connectors in production
- from scoping to monitoring
What does the Mistral API provide and why choose it for an AI project in France?
Mistral is a French AI publisher whose models are available by API for text generation, document analysis and calling your business tools. You choose Mistral when data sovereignty is a key criterion: inference can be localised in European datacentres, and open-weight models can be self-hosted on your own infrastructure for projects that cannot send data to a third-party cloud. It is the European alternative to US APIs for regulated sectors, public procurement and high-sensitivity projects.
What our clients build on the Mistral API
French business assistant, eu inference
Function calling into the ERP. Files stay with you: no regional Files API.
OCR 4.1 on invoices and credit notes
Blocks, labels, confidence, into Pennylane or the ERP. A document brick, not a generic LLM.
Support replies with Moderation 2
Classification, drafting, filter before display. Shieldstral / mistral-moderation depending on scoping.
RAG: index at the client, chat on the EU
Regional inference is not a regional control plane. Account and billing may stay outside the inference geography.
What this changes in your product
Engineering in service of a measurable outcome: enforceable EU inference, a readable cost, files on your side when needed.
Europe is no longer a misunderstanding
The global API and the European API are not the same thing. We set the right entry point when the contract requires EU inference.
The EU surcharge is readable in the quote
The compliance uplift shows clearly. You compare a number, not a vague enterprise promise.
Files stay on your side when needed
Chat on Europe, documents in your infra. We do not sell an EU guarantee incompatible with vendor storage.
Open-weight stays a plan B
Hosting the model yourself is an infra project, not an API key. We scope it separately, with the same business tests.
How we ship your Mistral connector
Scoping
Global or eu, models.list on the regional host, exclude Agents/Files/Batch if inference residency is a commitment.
Evals
French corpus, OCR on your documents, moderation before display. Dated model ID, not a -latest alias.
Development
server="eu" (SDK ≥ 2.70) or server_url, log hostname + model id + request id, token cap, internal LLM interface.
Monitoring
Deprecation table (Medium 3.1 retirement 31/08/2026, and others). Open-weight failover if the cloud contract breaks.
What the Mistral API allows
- Chat completions and function calling
- Function calling is the only regional tool. Business assistant into the ERP, without Agents API on the EU.
- OCR 4.1
- Blocks, labels, confidence. Invoices, CERFA, tenders. A document brick, distinct from the chat LLM.
- Voxtral Transcribe 2
- Batch STT, and Realtime live. Realtime / Files can leave the regional path: cross-check with the Transcription page.
- Regional inference 1.1×
- api.eu.mistral.ai, EU + EFTA datacenters. Control plane (account, keys, billing, analytics) may stay outside the geography.
Mistral API vocabulary
- api.eu.mistral.ai
- EU + EFTA inference host, 1.1× surcharge. Function calling only. No Agents, no Batch, no Files API. This is the page to open in a GDPR workshop.
- Control plane vs inference
- Account, keys, billing, analytics may still be handled outside the selected inference geography. Regional inference is not a regional control plane.
- EU hosting by default
- Help Center: by default, your data is hosted in the European Union. A distinct sentence from global inference, which commits to no location. Hold both, do not merge them.
- ZDR
- Zero Data Retention: another lever than the region (retention after the fact). Both may be needed. Enable it if the DPA allows, on top of regional.
- Open-weight
- Small 4, Large 3, Ministral 3 (3B/8B/14B) self-hostable. This is not the operated API. Infra project, same tests as the cloud.
- server="eu"
- SDK ≥ 2.70. Older: server_url. Log endpoint hostname, model id, request id: Mistral asks for it for audit.
The real constraints of the Mistral API
api.mistral.ai is not Europe
No inference location promised. A customer who hears Mistral equals EU on the global endpoint is wrong. api.eu.mistral.ai, or we do not say Europe.
Agents, Files, Batch off regional
A Files API RAG is not EU inference. Chat on api.eu, files with you. Otherwise the contractual commitment is false, and it must be said before coding.
The regional catalogue is a subset
models.list on api.eu before coding the ID. global mistral-large-latest ≠ available on EU. Fast deprecations, avoid -latest aliases.
Open-weight ≠ API key
Hosting Small 4 at the client is GPU infra and operations, not a URL change. The quote is not the same job as an api.eu API key.
api.mistral.ai or api.eu.mistral.ai?
Two hosts, two promises. The 1.1× and the exclusions (Agents, Batch, Files) are said before coding, not after the DPA.
| Criterion | api.mistral.aiGlobal | api.eu.mistral.aiEU inference |
|---|---|---|
| Inference location | None promised | EU + EFTA datacenters |
| Surcharge | List price | 1.1× tokens in/out, cache |
| Function calling | Yes | Yes, only regional tool |
| Agents / Batch / Files | Yes (global catalogue) | No |
| Models | Full catalogue | Subset, models.list |
| Control plane | Global | May stay outside the geography |
| The right case | Prototype, Agents features | Prod, opposable EU inference |
EU hosting by default (Help Center) and global inference with no commitment are two sentences. ZDR adds to the region, it does not replace it. Hostname logged on every request.
What we measure on a Mistral integration
The other AI building blocks
Mistral is combined more often than it is replaced. These options are discussed at scoping.
MistralWe build your Mistral connectorThis page
GeminiVertex eu, when Google multimodal is the right tool, not the lab.
HeyGenThe video avatar, when text does not carry the message.
ElevenLabsWe build your ElevenLabs connectorTranscriptionWe build your transcription connector
Anthropic (Claude)A model that calls your tools, not one more chat
OpenAIThe model writes into your tools, it does not chat
CursorThe agent reaches your internal tools, not just your codeWe combine Mistral with
The stack around Mistral on our projects.
Mistral: your questions
Three steps. Choose the host: api.eu.mistral.ai as soon as an EU inference commitment exists (SDK server="eu"), otherwise global and say so. models.list on that host, pin a dated ID. Chat and function calling on the EU, files with you (no regional Files API). Log hostname, model id, request id. Token cap, moderation before display. The sensitive part is not the chat call, it is not selling Agents/Batch/Files as EU inference, and not merging EU hosting by default with global inference.
It guarantees inference in EU + EFTA datacenters, at 1.1×, with function calling only. The control plane (account, keys, billing, analytics) may still be handled outside the selected inference geography. Agents, Batch and Files API are not on regional endpoints. ZDR is another lever. Help Center: data hosted in the EU by default, limited transfers, SCCs. We log the hostname and we write what the contract covers, no more.
At Mistral, EU inference is a dedicated host (api.eu.mistral.ai) at +10%, with a reduced catalogue. At Google, the Studio API has no region; residency goes through Vertex eu. Both require logging the endpoint. Mistral adds self-hostable open-weight and a European lab. Gemini adds Google multimodal and GCP. The choice is made at scoping (DPA, models, OCR, voice), not on a Europe label glued to the wrong URL.
Open-weight models (Small 4, Large 3, Ministral) yes: that is a GPU infra project, not an API key. The operated API (api.eu) is not on-prem. We plan the failover (same tokenizer, same tests) if the cloud contract breaks, without selling hosting as a simple URL change. Codestral stays more of an internal migration tool, off the customer page. The two paths can coexist: api.eu in production, open-weight as a documented exit.
A first chat flow on api.eu with function calling and hostname logs ships in two to three weeks. OCR 4.1, moderation, RAG (index with you) and an open-weight plan are closer to six to eight weeks. Duration depends on the catalogue actually served on api.eu (models.list) and the Agents/Files exclusions. We scope the perimeter up front and give you a firm estimate before we start, including the host.
A Mistral integration project?
Let's talk. 30 minutes to scope api.eu or global, what Agents/Files break, and tell you frankly what is opposable.
Discuss my Mistral project