Mistral for business: assistant, API, self-hosting or private deployment, and who is responsible for what

Four routes, and the question that separates them

Mistral describes itself as an AI lab that builds frontier and specialist open-weight models. Around those models it sells several things: Forge to customize models, Studio to build and govern agents, Vibe to bring the models into research, coding and everyday tasks, and the option to run the models on-premises, in your own VPC or on Mistral's cloud.

For a buyer, these become four routes: the assistant, the API with Studio, open-weight models you host yourself, and a private deployment with customization. The question that separates them is simple: who operates the infrastructure, and who handles your data? The rest of this guide answers it route by route. The responsibility split is our editorial reading of Mistral's pages, not a contract term.

Sources: mistral.ai · Mistral (usedby.ai)

The assistant: for people who do not want to build anything

Choose the assistant when the need is individual or team productivity without development. Mistral's pricing page lists Free at $0, Pro at $14.99 per month, Team at $24.99 per user per month, and Enterprise on request. Team adds up to 30 GB of storage per user, domain verification and data export. Enterprise lists custom models, agents and workflows, audit logs, SAML SSO and white label.

Here Mistral operates the service and you manage users and usage. The plan table lists an opt-out for model training, so check the setting before staff paste in sensitive material. A naming caveat: the current pages talk about Vibe, while the name Le Chat appears only in a 2025 Trust Center notice. The sources do not confirm that one replaced the other, so ask Mistral which product your quote covers.

Sources: Mistral (usedby.ai) · Pricing (mistral.ai) · mistral.ai

API and Studio: when models go inside your own product

Choose the API when a developer team will embed models in an application. Pricing is per million tokens, with input and output counted separately. The page gives Mistral Large at $0.5 in and $1.5 out per million tokens, a 50% reduction for batch processing, and up to 90% off cached input. OCR is priced per 1,000 pages, speech per minute, and tool APIs per call.

Mistral points to Medium for most tasks and coding, Small for cost-sensitive work, OCR for documents and Voxtral for audio. The Free plan includes access to Studio, which also has a dashboard tracking calls, endpoints and spending. You build and monitor the application; Mistral runs the models. The Trust Center lists Mistral Compute in France, Microsoft in Sweden and Norway, Google in the Netherlands, Belgium and the US for a US API endpoint, and CoreWeave for inference in the EEA.

Sources: Mistral (usedby.ai) · mistral.ai

Self-hosting and private deployment: control comes with the workload

For inference and compute, the homepage says you can run on-prem, in your VPC, or have Mistral run it, with inference kept in-region and dedicated GPU capacity reserved for predictable throughput. Use these routes when infrastructure control or data sovereignty outweighs convenience. Mistral's pricing page confirms its models can be self-hosted anywhere.

What you take on is the part the pages do not price: hardware, operations, scaling and security patching. One licensing point needs confirming. The pricing page says open-weight models such as Mistral 7B are Apache 2.0 licensed for research or individual use, while commercial deployments need a Mistral license with separate terms for derivatives and production use. Ask for the exact license per model before building on it. Enterprise private deployments are sold through sales, so no price is public.

Sources: mistral.ai · Mistral (usedby.ai)

Forge: when the model must learn your data

Forge is for companies that want a model adapted to their own data, use cases and requirements. Ericsson's page describes a partnership combining Mistral's model customization platform with Ericsson's R&D and network expertise, aimed first at automating legacy code translation and AI-assisted 6G research. Mistral's customer list also describes DSO National Laboratories as building a custom language model with Forge, and HTX's AI as built with Mistral Forge.

A related custom-model example is the Austrian Academy of Sciences' Apollo, a model for Ancient Greek trained on a corpus of 600 million words, built with Mistral and services partner Reply on European-based infrastructure. Mistral reports that work on damaged texts that once took years can now take hours. That is a vendor-reported outcome for one research project, not a benchmark, and the page does not present it as a Forge deployment. What the Forge cases do not settle is contractual ownership of an adapted model: the homepage's claim that what you create is yours is marketing copy, so confirm ownership terms in the contract.

Sources: Ericsson (mistral.ai) · Customers (mistral.ai) · mistral.ai · Austrian academy of sciences (mistral.ai)

Data and contract: who is controller, processor and trainer

Mistral's data processing addendum makes the customer the controller and Mistral the processor. It also lets Mistral act as controller to train its models unless the customer has opted out, to run automated moderation including abuse monitoring on APIs (except when zero data retention is activated), and to produce aggregated usage statistics. Feedback given through thumbs up or down, with the related input and output, is used for training.

After termination, personal data is deleted or returned, and the addendum says it is no longer accessible after 30 days. Subprocessor changes get notice, and a customer has ten days to object on data protection grounds. The Trust Center lists ISO 27001, ISO 27701 and SOC 2 Type II, plus a note that attackers briefly compromised non-core code repositories in May 2026, which Mistral says did not touch hosted services or user data. It also published a fix for Vibe shell permission vulnerabilities in September 2026.

Sources: Data processing addendum (legal.mistral.ai) · trust.mistral.ai · Mistral (usedby.ai)

What to confirm before signing

These are editorial suggestions, not Mistral capabilities. Ask which product name and contract cover your quote, whether training opt-out and zero data retention apply to your plan, which commercial license covers each open-weight model, who owns a customized model, what Enterprise and private deployments cost, and where each subprocessor processes your data.

A practical path: start with the Free plan, which includes Studio access, for a low-risk use case, move to self-hosting or a VPC only if control requires it, and consider Forge only when you hold domain data that general models handle poorly.

Sources: Mistral (usedby.ai) · Data processing addendum (legal.mistral.ai) · mistral.ai