Mistral AI Documentation
by Mistral AI
The official builder docs for Europe's open-weight frontier lab — agents, RAG, OCR and fine-tuning against models you can also self-host.
Overview
The documentation is organized around four products. Studio is the developer console and API platform and carries most of the engineering material: platform overview, 5-to-15-minute quickstarts, CLI and API-key setup, project scaffolding, SDK clients and the API reference. Models documents the current lineup, including the multimodal and agentic Mistral Medium tier, a Mistral Small release published under Apache-2.0 that can be self-hosted, the Voxtral audio-transcription models and a dedicated OCR service — which is the part enterprise RAG teams tend to care about most, since document parsing is usually the bottleneck rather than retrieval. Vibe covers the productivity and coding agent, and Admin covers organization, workspace and access management for teams. The tutorial track is task-shaped rather than reference-shaped: send your first API request, build an agent with tools, and build RAG with document search, then workflows and Skills. A companion cookbook repository (MIT-licensed, 2.3k stars, 1,100+ commits) extends this with notebooks on chat, embeddings, function calling, RAG, fine-tuning, evaluation, structured outputs, synthetic data generation, Pixtral image handling, OCR, text-to-SQL, moderation, agents, connectors, observability and custom MCP servers, plus integrations with LangChain, LlamaIndex, ChromaDB, Pinecone and thirty-odd other platforms. The docs carry a 2026 copyright and versioned model entries, so they track releases rather than drifting.
At a Glance
- Topic
- Models
- Level
- All Levels
- Format
- Documentation
- Cost
- Freemium
- Duration
- ~15 minutes for the first quickstart; ~4-6 hours across the agents, RAG, workflows and fine-tuning tutorials
- Provider
- Mistral AI
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Make your first Mistral API call and scaffold a project with the CLI and SDK clients
- ✓Build an agent that calls tools, and extend it with connectors and custom MCP servers
- ✓Implement RAG with document search over your own corpus using the documented tutorial
- ✓Use the dedicated OCR service to parse documents before they reach a retrieval pipeline
- ✓Fine-tune a Mistral model and evaluate the result against a baseline
- ✓Produce reliable structured outputs and apply moderation to model responses
- ✓Transcribe and work with audio through the Voxtral model family
- ✓Decide between the hosted API and self-hosting the Apache-2.0 open-weight releases
Highlights
- •Open-weight releases under Apache-2.0 sit alongside the hosted API, so the same tutorial code path works for a self-hosted deployment — rare among frontier labs.
- •First-class OCR and document-parsing service documented as a product, not an afterthought, which is the real bottleneck in most enterprise RAG builds.
- •Task-shaped tutorials (first request, agent with tools, RAG with document search) rather than a wall of endpoint reference.
- •A separate MIT-licensed cookbook repository with 2.3k stars extends the docs into fine-tuning, evaluation, synthetic data, text-to-SQL and custom MCP servers.
- •Documents an EU-headquartered provider with its own console and admin/workspace controls — the practical answer when data residency rules out US-hosted inference.
- •Model entries are versioned and the docs carry a 2026 copyright, so pages track the current lineup instead of describing a two-year-old release.
Who It’s For
Best For
- ✓AI engineers evaluating an alternative to OpenAI and Anthropic for production workloads
- ✓Teams with EU data-residency constraints who need a documented non-US provider
- ✓Developers who want open weights they can self-host with a hosted API for prototyping
- ✓Engineers building document-heavy RAG pipelines that need real OCR before retrieval
Prerequisites
- •Basic Python or JavaScript and comfort calling a REST API
- •A Mistral account and API key; usage beyond the free tier is billed
- •Conceptual familiarity with embeddings and RAG before the document-search tutorial
FAQ
What is Mistral AI Documentation?
Official Mistral AI documentation for developers building on its model family and Studio platform. It matters to AI engineers specifically because Mistral ships open-weight models under Apache-2.0 alongside its hosted API, so the same tutorials apply whether you call the endpoint or run the weights yourself. Coverage spans quickstarts, agent building with tools, document-search RAG, workflows, Skills, fine-tuning, OCR and audio, plus SDK clients and a full API reference.
Is Mistral AI Documentation free?
Mistral AI Documentation offers free content, with paid options for certificates or premium features.
What level is Mistral AI Documentation for?
Mistral AI Documentation is aimed at a all levels audience. Recommended background: Basic Python or JavaScript and comfort calling a REST API, A Mistral account and API key; usage beyond the free tier is billed, Conceptual familiarity with embeddings and RAG before the document-search tutorial.
How long does Mistral AI Documentation take?
Expect roughly ~15 minutes for the first quickstart; ~4-6 hours across the agents, RAG, workflows and fine-tuning tutorials. Most learners work through it at their own pace.
What will I learn from Mistral AI Documentation?
You'll learn: Make your first Mistral API call and scaffold a project with the CLI and SDK clients; Build an agent that calls tools, and extend it with connectors and custom MCP servers; Implement RAG with document search over your own corpus using the documented tutorial; Use the dedicated OCR service to parse documents before they reach a retrieval pipeline; Fine-tune a Mistral model and evaluate the result against a baseline; Produce reliable structured outputs and apply moderation to model responses; Transcribe and work with audio through the Voxtral model family; Decide between the hosted API and self-hosting the Apache-2.0 open-weight releases.
Topics
Sources
This page was written from 2 sources, 1 on domains other than docs.mistral.ai.