Mistral for business: Vibe (formerly Le Chat), agents and sovereign deployment (2026 guide)

Mistral AI is a French company that builds AI models, the programs that understand and write text, and sells them in two forms: an assistant for teams, Mistral Vibe, known until 2026 as Le Chat, and models that companies install or integrate into their own tools.

Sovereignty, applied to AI, means an organization’s ability to keep control of its data: knowing where it is processed, who can access it and under which law. It depends on how the tool is deployed, on the contract and on the training setting, and the vendor’s nationality is only one part of it.

Mistral offers five ways to deploy, from an online subscription to a model installed on your premises, more than the three other major assistants compared below. This guide, current as of October 14, 2026, gathers facts checked in Mistral’s documentation and in official texts, along with a method for choosing.

The essentials in six points

• Le Chat is now called Mistral Vibe: same address, same accounts, with one mode for office work and one for code. For a safe first use, go straight to the practical part.
• Individual plans can be used to train Mistral’s models unless the user turns it off. The Enterprise plan excludes it by default.
• A French vendor is not enough: what counts is where the computing runs, who can access it and what the contract says.
• Mistral can be deployed in five ways, from an online subscription to a model installed on your own servers. Each protects more or less, and takes more or less effort.
• The French State chose Mistral for its own assistant, L’Assistant, hosted by Outscale on a SecNumCloud-qualified cloud and rolled out to State employees in June 2026.
• Several Mistral models are free for commercial use and install without a contract: Large 3, Small 4 and the Ministral 3 family, under the Apache 2.0 license.

What Mistral sells in 2026

Mistral sells three things, which need to be told apart before any decision.

• Vibe, the assistant for teams, available in the browser, as a mobile app and, for developers, on the command line.
• Models through the API, on the Mistral AI Studio platform. An API is a technical access point through which software sends a request to the model and gets the answer back, with no interface in between.
• Open-weight models, to download and run on your own machines. They are covered further down, in the technical part.

What Vibe does for an employee

The Vibe product page mentions research, document creation and meeting preparation. An example, built from the features described in Mistral’s announcement of May 28, 2026:

• the request: “Prepare my Thursday meeting with supplier X from our latest exchanges and the current contract”;
• what Vibe reads: the related emails in Outlook and the contract in SharePoint, through its connectors;
• what it returns: a briefing note in Canvas, its drafting space, with open issues and questions to raise, exportable to SharePoint.

The same request can be scheduled to run every week. The employee reviews the note before using it.

Public references

• The French Ministry of the Armed Forces awarded Mistral a framework agreement on January 8, 2026, according to Next, open to the armed forces, their directorates and several public bodies.
• DINUM, the French government’s digital agency, built “L’Assistant” with Mistral, a conversational tool hosted by Outscale on a SecNumCloud-qualified cloud. The press release of June 16, 2026 announces its rollout to all State employees, after ten months of trials with 10,000 of them.
• Microsoft announced on July 21, 2026 that Mistral Medium 3.5 was coming to Microsoft Foundry and Copilot Studio.

From Le Chat to Vibe

Mistral announced on May 28, 2026 that “Le Chat is now Vibe”, with “one agent and one licence across work and code”. Its help center dates the effective switch to August 12, 2026.

Nothing was lost along the way: accounts, plans, conversations and saved content carried over, at the same address, chat.mistral.ai. Vibe has three modes.

• Vibe Work, the work mode: you hand it a multi-step task and it carries it out across your tools, email, calendar, documents.
• Vibe Code, for developers, on the command line, in the VS Code editor or on the web.
• Vibe Chat, the turn-based conversation Le Chat offered.

Le Chat is still the name most people type into a search engine. Both refer to the same service.

Plans and pricing

According to Mistral’s pricing page, checked on October 14, 2026, prices are shown in dollars, excluding tax.

Plan	Price	What it adds
Free	Free	Limited messages and searches, 100+ connectors
Pro	$14.99 per month	Higher limits, code, spreadsheets and slides, $15 of API credits per month
Team	$24.99 per user per month, $50 minimum	Domain verification, knowledge connectors, 30 GB storage per user, data export
Enterprise	On request	Custom models, agents and workflows, audit logs, SAML SSO, white label, private deployments

A student rate of $5.99 per month is available for up to twelve months.

Three Enterprise terms deserve a word. SSO (single sign-on) lets employees log in with their company account; SAML is the standard behind it. Audit logs record who did what. White label lets the organization present the tool under its own name.

The choice between Team and Enterprise comes down to three needs: SSO, audit logs and private deployment. As of October 14, 2026, none of the three is part of the Team plan.

Cost at scale

At list price, 2,000 users on the Team plan cost 2,000 × $24.99 = $49,980 per month excluding tax, about $600,000 a year. Enterprise is negotiated on request, and its price is not public.

For an on-premises deployment or self-hosted models, the cost is driven mostly by hardware, and therefore by the chosen model and the volume of requests. We give no figure without knowing those two parameters.

Agents, connectors and MCP

An agent is an assistant that does more than answer: it reads documents, works through steps and acts in software. In Vibe, an agent combines instructions, document libraries and tools.

A connector is a ready-made link between Vibe and a piece of software. The Vibe product page lists more than 100 tools, including email, calendar, Slack, GitHub and Jira, “through connectors and full MCP compatibility”.

MCP, the Model Context Protocol, is an open standard that lets an AI assistant plug into software in a uniform way. A tool that exposes an MCP server can be used by Vibe as well as by other compatible assistants.

In its May 28 announcement, Mistral mentions connections to Google Workspace, Outlook, SharePoint, Slack and GitHub, scheduled tasks on a daily, weekly or monthly cadence, and permissions set by the administrator.

An agent sees whatever its connectors open up. Plugging in an entire mailbox or an overshared SharePoint gives the agent far more than it needs. The first setting to get right is the scope of access, before choosing tasks.

Where to start

If you are trying Mistral for your team, with no technical background, this is the order we recommend.

1. Check the training setting. On an individual account, open the privacy settings and turn off data sharing for training, as Mistral’s help center explains.
2. Start with public or non-sensitive internal documents: a procedure, a brochure, a summary note. Nothing personal, nothing covered by professional secrecy.
3. Build a first simple agent: clear instructions, a library of a few documents, and a request to cite the source of each answer.
4. Test it with the team’s real questions, over a period set in advance, and log each mistake with the question that caused it.
5. Ask IT before plugging in a connector to email, calendar or a business application.

If your use stops there, a properly configured Pro or Team subscription is enough, and you can stop reading here. What follows is for IT departments, CISOs (chief information security officers), DPOs (data protection officers), legal teams and public buyers.

Models and their licenses

The weights are the billions of parameters that make up a model; once they are published, the model can run on any sufficiently powerful machine. The model overview in Mistral’s documentation, checked on October 14, 2026, gives each model’s license. An Apache 2.0 license allows use, modification and commercial use, royalty-free.

Model	Use	License
Mistral Large 3	General purpose, text and image	Apache 2.0
Mistral Small 4	Compact general purpose, reasoning and code	Apache 2.0
Ministral 3 (3B, 8B, 14B)	Small models for a workstation or modest server	Apache 2.0
Mistral Medium 3.5	Agents and code	Modified MIT
Mistral Large 4	General purpose, text and image	Announced as open, to be read when weights ship

Mistral Large 4, announced on October 6, 2026, has about 1 trillion parameters, 49 billion of them active for each computation, according to Simon Willison, who reports Mistral’s promise to release the weights at “the end of this month”. Guillaume Lample, Mistral’s co-founder, says it was trained on 3,800 GPUs in Mistral’s European data centers.

Its license has not been published yet. Until it is, “open weights” does not guarantee free commercial use: the rule applies to any model.

Five ways to deploy

The terms in the table

• The cloud means servers rented from a provider and used remotely; a private cloud is infrastructure reserved for one organization, at a host of its choosing. An on-premises deployment puts the service in the organization’s own data centers.
• Data residency is a contractual commitment that data stays stored in a given area, for example the European Union.
• Subprocessors are the providers a vendor relies on for part of the processing, starting with server hosting.
• Exposure to non-European law means that an authority from a country outside the EU can, under its own law, compel a provider to hand over data.
• SecNumCloud is the security qualification that ANSSI, the French national cybersecurity agency, issues to hosting providers; among other things, it requires protection against such laws.

The table

Option	Where the data is	Exposure to non-European law	SecNumCloud possible	Training	SSO and audit	Effort
Vibe Pro or Team, online	EU by default, at Mistral	Depends on the subprocessors Mistral lists, to be read	No	Pro: on by default, turn it off. Team: administrator can turn it off	No	Low
Vibe Enterprise, Mistral Cloud	Mistral Cloud, with data residency	Depends on the contract and subprocessors	No qualification published as of Oct 14	Off by default	Yes	Low
Vibe Enterprise, private cloud	At the chosen host	That of the host	Yes, with a qualified host	Off by default	Yes	Medium
Vibe Enterprise, on-premises	On your servers	None through hosting	Not applicable, it is your infrastructure	Data does not leave; the contract governs vendor access	Yes	High
Self-hosted open models	On your machines	None through hosting	Not applicable, or with a qualified host	Nothing sent to Mistral	To be built	High

What the table means

Online subscriptions. Mistral’s help center states: “By default, your data is hosted in the European Union.” It adds that data may be transferred temporarily outside the EU to the subprocessors listed in its trust center. Read the subprocessor list before signing, because the infrastructure provider matters as much as the vendor.

The Enterprise plan. The product page puts it plainly: “Enterprise customers can deploy Vibe on-premises, in a private cloud, or on Mistral Cloud with full data residency.”

The qualified private cloud. Outscale, a Dassault Systèmes subsidiary, offers Mistral AI Studio and Vibe on its infrastructure, certified HDS, the French certification required to host health data. Its public cloud received SecNumCloud 3.2 qualification in December 2023. The qualification covers the hosting infrastructure, not the application service running on it, whether Vibe or the French State’s L’Assistant.

Why there is no exposure on-premises

Foreign law is exercised over a provider that holds the data. When the data stays on your servers, no provider holds it. Exposure only comes back through the access you open: the vendor’s remote maintenance, telemetry, support. An on-premises contract must therefore state whether the vendor accesses the system, when, and what is sent back to it.

Self-hosted open models follow the same logic. Installing Mistral Small 4 or a Ministral 3 on an internal server requires no contract with Mistral. Nothing leaves your premises. In return, the interface, access control, logging and updates have to be built, as our article on sovereign AI and local data explains.

Mistral’s security commitments

According to Mistral’s help center, dated August 12, 2026, Mistral complies with SOC 2 Type II, an independent US audit that checks security controls work over time, and ISO 27001 and 27701, the international standards for information security and privacy management.

The full SOC 2 report is reserved for Enterprise customers. We found no published HDS certification for Mistral’s own cloud.

What the law says

Texts binding on the French State

Three texts govern how the French State administration uses the cloud.

• The “cloud au centre” doctrine (cloud first), set by a circular of July 5, 2021 and updated by circular no. 6404/SG of May 31, 2023. It requires data of particular sensitivity to be hosted on a SecNumCloud-qualified offering, or an equivalent European qualification, protected against non-EU regulations.
• Article 31 of law no. 2024-449 of May 21, 2024, known as the SREN law. For such data, it requires a provider that protects against any access by third-country authorities not authorized by EU or Member State law.
• Decree no. 2026-272 of April 14, 2026, which implements that article. It refers to an ANSSI security framework, requires reversibility clauses and provides time-limited exemptions for projects already under way.

Who is bound

The decree covers State administrations, State operators and the public interest groupings it lists. A French local authority is not bound by it. A public body should check whether it counts among the State’s operators.

Sector rules still apply to everyone: hosting health data on behalf of a third party requires an HDS-certified host, under Article L. 1111-8 of the French Public Health Code.

Article 31 covers two categories of data: data covered by a secret protected by law, and data needed for the State’s essential missions, including national security, public order, and people’s health and lives. A company or local authority handling comparable data can apply the same framework, without being required to.

US law

The CLOUD Act, a 2018 US law on government access to data held by providers, wrote a rule into Title 18, Section 2713 of the US Code.

A provider subject to US law must disclose data “within such provider’s possession, custody, or control, regardless of whether such communication, record, or other information is located within or outside of the United States.”

The Data Privacy Framework is the data protection arrangement between the EU and the United States. Since the European Commission’s decision of July 10, 2023, it has made transfers to participating US companies lawful under the GDPR.

The EU General Court upheld it on September 3, 2025 (case T-553/23); an appeal lodged on October 31, 2025 is pending before the Court of Justice (case C-703/25 P).

That framework settles whether a transfer is lawful. It does not remove a US provider from the reach of its own national law, and therefore does not, on its own, meet the protection requirement set by Article 31.

For Mistral, the lesson is twofold. A French vendor reduces that exposure without eliminating it: its hosts and subprocessors count too. And a Mistral model served from qualified infrastructure, or installed on your premises, reaches a level of protection that the online subscription does not.

Before signing: three checks

• Reversibility and portability: the contract should provide for returning data, agents and logs in a usable format, and a transition period if you leave.
• Impact assessment with the DPO: an assistant connected to email or personal files processes personal data at scale; the DPO decides whether a data protection impact assessment is required.
• AI Act: an organization using an AI system under its own authority is its deployer under the EU AI regulation, with obligations of its own. Our article on what applies since August 2, 2026 covers them.

For public buyers

As of October 14, 2026, we found no public listing of Vibe with UGAP, the French State’s central purchasing body, nor any framework agreement open to local authorities; the Armed Forces agreement is reserved for the ministry and the bodies it designates. The purchasing route should therefore be checked with UGAP, your regional purchasing body or Mistral, failing a contract awarded on your own.

Mistral compared with Claude, ChatGPT and Copilot

We train on all four tools. The criterion that separates them most clearly is the deployment option.

To our knowledge as of October 14, 2026, Mistral is the only one of the four to offer its assistant for on-premises deployment. ChatGPT Enterprise, Claude Enterprise and Microsoft Copilot are delivered online by their vendors.

Since August 6, 2026, Anthropic has let teams run Claude Code sessions on their own machines, but the conversation and the model’s computation stay at Anthropic.

OpenAI has published open models, gpt-oss, under the Apache 2.0 license since August 2025, but not ChatGPT itself.

Tool	Strengths	Limits to know
Mistral Vibe	French vendor, five deployment options including on-premises, open models under Apache 2.0, used by the French State	Training on by default for individual plans, no published SecNumCloud qualification for its own cloud
Claude (Anthropic)	Long work across folders of files with Cowork, skills shared across the organization	US vendor, assistant delivered online
ChatGPT (OpenAI)	Agent mode, projects, shared agents on Business and Enterprise plans	US vendor, assistant delivered online
Microsoft Copilot	Inherits Microsoft 365 permissions, Purview labels and existing governance	US vendor; agent costs in credits to be managed

The four are not mutually exclusive. Mistral Medium 3.5 has in fact been offered in Copilot Studio since Microsoft’s announcement of July 21, 2026. One possible architecture keeps a US assistant for everyday use and reserves Mistral, on a qualified cloud or on-premises, for sensitive data.

To go deeper on each:

• our Claude Cowork guide and our guide to deploying Claude in the enterprise;
• our Copilot Studio guide, on Microsoft’s agents, their credits and their governance.

Training your teams

The deployment option sets what the tool can do with the data. Usage sets what employees hand to it. Three situations need preventing: a client document pasted into a personal account, a connector opened to the whole mailbox, an answer reused without checking its source.

Training therefore covers three habits: knowing which data goes into which tool, scoping an agent, and checking what it produces. These habits apply to Vibe as much as to any other assistant.

Going further

Colombani.ai is certified Qualiopi, the French quality certification for training providers, which is a condition for public and pooled funding.

• For business teams and executives, the Mistral training comes as a half-day briefing for executives and a day on Vibe, its agents and connectors.
• To choose a deployment option, subscription, qualified cloud, on-premises or open models, Colombani.ai runs a sovereign architecture scoping engagement.
• For technical teams, the Local AI course has participants install Mistral’s open models on your own machines, with no data leaving.

For French companies, these programmes are eligible for funding from OPCOs, the skills operators that fund private-sector employee training. For a local authority or public body, training is paid from the employer’s training budget; Colombani.ai provides a quote, programme and agreement that meet public buyers’ requirements.

Colombani.ai’s founder, Ulysse Trin, holds the Claude Certified Architect Professional, issued by Anthropic, and nine practitioner certifications issued by OpenAI.