In a few years, Mistral AI has gone from a range of models to a set of building blocks covering generation, reasoning, multimodality, code, agents, document intelligence, search, connectors, customization, and deployment. For a European company, this offering is attractive because of its diversity and the options for control or hosting.
The difficulty is avoiding two opposite mistakes: reducing Mistral to "a French model" or adopting its entire ecosystem without benchmarking or output architecture. A stack must be composed based on the use case, data, and operational constraints.
Models, licenses, preview statuses, prices, regions, and performance evolve rapidly. Any decision must be revalidated in official sources and through internal testing.
Mapping the bricks available in 2026
Mistral Small 4
Announced in March 2026, Mistral Small 4 combines conversation, reasoning, coding, and image understanding in a hybrid model. Mistral features a mixture-of-experts architecture with 119 billion parameters in total, 6 billion active per token outside of input/output layers, a context window of 256k, and an Apache 2.0 license.
This combination is aimed at versatile uses and open deployment. The performance figures published by the provider must be reproduced on the target tasks and runtimes.
Mistral Medium 3.5
Presented in May 2026 in public preview, Medium 3.5 is described as a dense model with 128 billion parameters, with 256k context, reasoning, code, vision, and agentic task capabilities. Mistral announces open weights under a modified MIT license and possible self-hosting from four GPUs depending on the configuration.
The preview status implies checking stability, support, licensing, and behavior before critical production.
Forge
Forge aims at building business models anchored in proprietary knowledge. This approach goes beyond standard fine-tuning and can involve organizations that have data, a strategic case, and sufficient evaluation capacity.
It does not automatically replace RAG. It is necessary to determine what must be learned in behavior and what should remain retrievable, citable, and updatable.
Vibe and the agents
Vibe covers multi-step coding and work uses, with local or remote sessions depending on the offerings. The agentic mode allows manipulating tools and files during a long task.
Adoption in a company requires sandbox, permissions, review, secrets, costs, and integration into the delivery cycle. The tool is not an exception to development rules.
Mistral OCR 4
The OCR 4 family targets document understanding: text, structure, tables, layout, and visual content. It can feed a document intelligence pipeline or a RAG.
Quality must be tested on real documents: scans, forms, tables, languages, handwriting, rotation, and low quality. The best average score does not guarantee exact accounting extraction.
Search Toolkit
Presented in public preview in May 2026, Search Toolkit offers components for ingestion, retrieval, and evaluation. It can reduce the plumbing of a RAG chain, but its data model, portability, and integration with rights need to be examined.
Studio, connectors and MCP
Mistral announced integrated and customized connectors, centralized for multiple applications and accessible via API or SDK. MCP facilitates the exposure of tools and data, but each connector must be registered, authenticated, limited, and audited.
Infrastructure and regional inference
In August 2026, Mistral presented new regional inference capabilities, open models, and a European infrastructure roadmap. For a sovereign project, check the actual region of each component, the subcontractors, support, backups, and transfers, rather than just the supplier's nationality.
Map of the bricks of a Mistral stack connecting use cases, models, document intelligence, search, agents, connectors, customization, and deployment.
Choose the model from the task
An architecture can use several models:
- compact and fast for classification;
- more capable for complex synthesis;
- OCR for ingestion;
- code template for development;
- local model for sensitive data;
- API for advanced peaks or tasks.
Routing depends on the expected quality, cost, latency, and data. A gateway hides the differences in APIs while retaining the explicit specific capabilities.
Do not choose Small or Medium from the name. Build the evaluation set and then measure.
Typical architecture of a documentary RAG
A possible stack:
- documentary deposit;
- OCR and parsing;
- standardization;
- segmentation;
- embeddings ;
- index ;
- research and reranking;
- generation model;
- quotations;
- evaluation;
- observability.
Mistral can provide several components, but the company can keep its vector database, its search engine, or its observability. This modularity facilitates reversibility.
The rights of the document are propagated to the output. The OCR must not flatten the permissions.
Build a document intelligence chain
For an invoice or a file, separate:
- file ingestion and control;
- classification;
- OCR/layout ;
- structured extraction;
- schema validation;
- business rules;
- reconciliation;
- human review;
- archiving.
The model can suggest the fields, but the totals, dates, and critical identifiers are validated by rules and human sampling.
Measure by field and by type of document. An overall rate hides errors in the amounts.
Deploy a code agent
Vibe or another agent reads the repository, modifies files, and runs tests. The company provides:
- deposit instructions;
- reproducible environment;
- isolated branch;
- tests ;
- static analysis;
- protected secrets;
- limited network;
- mandatory review;
- quality metrics.
Compare the agent to a baseline on real tickets: time until PR accepted, regressions, review effort, and debt created.
Determine if Forge is justified
A deep personalization is relevant when:
- the capacity is strategic;
- the data are unique and authorized;
- The behavior is not resolved by prompt/RAG;
- The volume justifies the investment;
- The team knows how to evaluate;
- Governance and hosting are defined.
Start with a baseline, a RAG, and possibly a targeted fine-tuning. A custom model without a benchmark test can produce an expensive asset that is impossible to evaluate.
Evaluate licenses and statuses
"Open weights" does not always mean Apache 2.0. Each model has its own license, restrictions, and conditions. Check:
- commercial use;
- modification;
- redistribution;
- derivatives;
- sectors;
- attribution;
- responsibility;
- support.
The public preview status indicates a risk of change. Plan for compatibility contracts and a fallback.
Sizing self-hosting
The number of parameters is not enough. Consider:
- precision and quantification;
- memory for the weights;
- KV cache;
- context;
- competition;
- flow;
- time-to-first-token ;
- GPU and interconnection;
- energy;
- high availability;
- exploitation.
Performance announcements on a given number of GPUs are a starting point, not production sizing. Benchmark on the actual hardware, runtime, and prompts.
Use PHP Mistral Client without coupling the business logic
Partitech maintains a PHP library dedicated to Mistral. In a Symfony or PHP application, it can facilitate access to the APIs, but the business code must depend on internal interfaces: generation, embeddings, tools, streaming, and errors.
This abstraction allows to:
- test;
- change the model;
- router ;
- add metrics;
- centralize retries and security;
- adopt a new version without rewriting the domain.
Specific features remain exposed through declared capabilities rather than through the lowest common denominator.
Governing the connectors
An MCP connector to CRM, files, or knowledge base is access to the IS. Save:
- owner;
- data;
- tools;
- permissions;
- transport;
- authentication;
- logs ;
- version ;
- environment;
- review date.
Write actions require confirmations and idempotence. The model never receives global rights for convenience.
Build the benchmark
For each task, compare at least:
- Mistral candidate model;
- current baseline;
- other relevant model;
- possibly self-hosted version.
Measure quality, French, format, tools, citations, safety, latency, throughput, cost, and operation. Check adversarial cases and long documents.
The benchmarks published by Mistral describe the potential of the model, not the guaranteed performance of the application.
Plan an output architecture
Keep:
- prompts ;
- assessment games;
- data;
- index in an exportable format;
- internal contracts;
- traces;
- configurations;
- fine-tuning artifacts according to rights.
An exit plan can replace a model without losing knowledge or evaluation.
Compose, measure, then industrialize
The Mistral 2026 ecosystem offers an interesting range for European projects, open, documentary, and agent-based. The right stack is not the one that uses the most products, but the one that satisfies the task with the fewest unjustified dependencies.
Partitech can carry out benchmarks, integrate APIs via PHP Mistral Client, deploy models, build the RAG, the agents, and document intelligence, then implement security and observability. The decision remains based on project data, not on a product sheet.
Let's talk about your project
Benchmark and integrate a Mistral stack adapted to your constraints with Partitech. Contact Partitech.