Colombani.ai's Mistral AI training teaches your teams to work with Vibe, the artificial intelligence assistant from the French vendor Mistral AI, which replaced the Le Chat assistant in May 2026. Vibe does more than answer a question: it takes on a complete piece of work, a piece of research, a memo or a spreadsheet analysis, inside your organisation's tools. Your leadership also settles the settings that protect your data.
From €1,000 excl. VAT per half-day and €1,800 per day for an in-company group; final price depends on group size and length, on quote.
Updated on
To download: Sovereign deployment selection grid →
Colombani.ai's Mistral AI training, from a Qualiopi-certified provider (certificate QUA009472), comprises three programmes: the executive briefing, a four-hour half day for leadership; one day for business teams on Vibe Work, the mode in which Vibe carries out a complete piece of work; one day of local AI to run open models on your own machines. Each session starts from the plan chosen, Pro, Team or Enterprise, and has the use of your data for model training settled before the first document is uploaded. From €1,000 excl. VAT per half-day and €1,800 per day for an in-company group; final price depends on group size and length, on quote. OPCO funding is possible.
Mistral announced on 28 May 2026: "Le Chat is now Vibe", meaning one agent and one licence for work and code, with conversations, settings and subscriptions carried over (source). A second step is under way: the Chat tab, which kept the Le Chat experience, is being removed progressively and Chat mode is being folded into Work (source). For an organisation, this affects the agents it has built, the memories it has accumulated and its teams' habits.
Vibe brings together Work, the agent for long tasks on web and mobile, and Code, for developers, in the terminal, in VS Code or in remote sessions. Mistral announcement
Accounts migrate progressively: a single conversation list, the choice between quick answers (Fast) and deeper reasoning (Think) in the message box, Skills in place of agents. Release notes
Enterprise administrators choose their migration date within a window of about six months; the Chat tab stays visible during that period. Release notes
After about six months, every organisation works in Work. Skills replace agents: the instructions of agents still in use need to have been carried over into Skills before then. source
| In Le Chat (Chat mode) | In Vibe Work | What the organisation must do |
|---|---|---|
| Agents | Skills: a SKILL.md file of instructions, templates and examples, loaded only when the task matches its description | List the agents in use, carry their instructions over into Skills and decide which will be shared across the whole workspace |
| Memories | Knowledge Base: pages you can browse and edit, recalled only when useful; an indicator appears each time it is read or changed | Have each person review what has migrated, and set the rule for sensitive data |
| Deep research | The deep-research Skill, which cites its sources | Learn to call it and to check the sources it cites |
| Think mode | The Fast or Think choice in the message box | Show when deeper reasoning is worth it |
| Simple conversation | Work answers short requests directly and opens a task list for long ones; a temporary chat, kept out of history, remains possible | Change habits: a request describes the expected result instead of being a single isolated question |
Mapping based on the release notes and the documentation for Skills and the Knowledge Base, read on 30 October 2026. According to Mistral's documentation, Vibe's Skills follow the open Agent Skills standard, originally designed by Anthropic: a Skill written well for Vibe carries over to the other tools that adopt this standard.
For an Enterprise customer: four decisions before the switch-over date
What we settle with you: the migration timetable, the inventory of agents and the writing of the Skills that replace them during the course, on your own cases.
Qualiopi-certified provider
Certificate QUA009472 for training actions: funding can be requested from your OPCO or the FIF-PL.
See the certificate ↗Overall satisfaction: 4 out of 4
Six sessions and fourteen participants as at 23 September 2026, across all programmes: 96% of objectives achieved, 100% recommendation, no dropouts.
See how it is measured →Local AI, the first course delivered
The first training session of Colombani.ai, a Qualiopi-certified provider, covered AI installed on workstations: all of its learning objectives were achieved.
See the Local AI training →On-site, anywhere in France
At your premises or remotely, always in your Vibe workspace or on your machines, for one to ten participants per group.
Training across France →Each programme is part of Colombani.ai's Qualiopi catalogue and builds on your documents and processes: in your Vibe workspace for the first two, on your workstations or servers for the third. Objectives and assessment are those of the catalogue; the materials and installation guide are written for Vibe.
What Vibe changes in each of your departments, shown live on your cases; what a European vendor does and does not bring: hosting, sub-processors, training on data, deployment options; choosing between Team and Enterprise, the GDPR and EU AI Act framework.
What you receive
Your teams learn to hand complete pieces of work to Vibe Work without writing code: research and drafting in Canvas, spreadsheet analysis, Skills for recurring tasks, the Knowledge Base, connectors to email and documents, scheduled tasks.
What you receive
Running an open model, Mistral or another, on your workstation or server, chosen to fit the hardware available, then querying it on your documents through a graphical interface: files stay on your machines and no request goes to a vendor.
What you receive
The seven sequences of the "Working with AI agents" programme, with the lengths set in the programme, and what participants do in Vibe during each. The learning objectives do not change from one tool to another.
What distinguishes an answer from the former Chat mode from a task carried out by Work: visible plan, tool calls on display, clarifying questions.
Document libraries, drafting in Canvas, analysing a spreadsheet and returning an Excel file, project folders.
Describing the expected result, the sources and the checks; writing the description that triggers the Skill at the right moment.
Linking Vibe to email, calendar and shared documents, telling read functions from write functions, running a web search.
Creating a Skill for one's role, sharing it with the workspace, scheduling a weekly task, adding a custom MCP connector when the business tool requires it.
Training setting, approval before sending or deleting, Knowledge Base and sensitive data, choosing between the cloud and local AI by confidentiality level.
Assessment of what has been learned, individual action plan and satisfaction survey.
Sovereign architecture consulting
To design an Enterprise deployment in a private cloud or on-premises, or open models on a server: architecture, access, logging, a costed exit plan.
See the consulting offer →For the legal department and the data protection officer: AI system mapping, risk register, impact assessment template and a compliance plan ranked by priority.
See the programme →For lawyers, notaries and accountants: professional secrecy, what may be entrusted to an online tool and what calls for a model installed in-house.
See the programme →Beyond ten participants, training becomes a programme, scoped on quote with the IT department and the business units.
One department or site at a time, each on its own cases, in the order agreed with the Vibe administrator and the migration timetable.
One or two champions trained in each department, who spread good practice, maintain the shared Skills and report difficulties.
Colombani.ai has already trained trainers on local AI and Mistral models: choosing the model for the hardware, sovereignty, a prescription grid. The same approach prepares your trainers for Vibe.
Indicators agreed at scoping, reviewed after each wave with the administrator, to adjust the programme rather than repeat it.
List prices shown by Mistral on 30 October 2026 on its pricing page, in US dollars and excluding taxes, as the page shows them by default. Subscriptions are paid by your organisation.
| Plan | List price excl. tax | For whom | What it includes | What it does not include |
|---|---|---|---|---|
| Free | $0 | An individual trying it out | Limited messages and searches, more than a hundred connectors, limited coding sessions | Conversations are used for training unless the user opts out: a plan to rule out for any work document |
| Pro | $14.99 per month | One user, a freelancer, a team trying it out | Higher limits, enough to code all day in the terminal, the editor or on the web, spreadsheets and slides, monthly API credits | No managed workspace; training on by default, to be turned off by each user |
| Team | $24.99 per user per month | A team or a department | Shared workspace, up to 30 GB of storage per user, email domain verification, data export | No single sign-on through the company directory (SAML SSO), audit logs, Admin API or private deployment |
| Enterprise | On quote | An organisation that must frame and trace usage, and choose where its data is hosted | SAML SSO, audit logs, Admin API, white label, custom models, agents and workflows, self-hosted, private-cloud or on-premises deployment, custom service-level commitments | No public price; no public documentation of on-premises requirements |
| Open models | No paid licence, with one exception | A technical team running the models on its own machines | Mistral Small 4, Mistral Large 3 and Ministral 3 under the Apache 2.0 licence; Mistral Medium 3.5 under a modified MIT licence; Mistral Large 4 under the Mistral Ecosystem Open-Weight licence, which Mistral also describes as a modified MIT licence | No interface, connectors or maintenance: everything has to be built; for Large 4, a separate commercial licence is required from anyone reselling access to the model as a service |
At list price, 100 users on Team come to $2,499 excluding tax per month. The amount of API credits included in Pro is not given here: Mistral's page shows two values depending on the plan variant, to be checked at purchase. Licences read on the models page and in the announcement of 29 October 2026. The Mistral for business guide compares the five deployment modes.
Mistral AI is a French company, and its customers' data is hosted in the European Union by default. That is not enough to rule out all non-European law, since some of the sub-processors Mistral lists may process data outside the EU, and it does not settle the questions the GDPR raises: who processes the data, where it travels, what it is used for. The table separates what a third party attests, what Mistral states and what depends on your settings.
| Topic | What is established as at 30 October 2026 | Nature |
|---|---|---|
| Hosting location | European Union by default; United States if the US API endpoint is chosen explicitly (source) | Stated by Mistral |
| Transfers outside the EU | Possible, depending on the feature, to the sub-processors listed in the Trust Center, covered by the safeguards of GDPR Article 46; on Enterprise, some features involving them can be switched off for the organisation (source) | Stated by Mistral, adjustable on Enterprise |
| Where API computing runs | Europe or the United States, by choice, since 11 August 2026, subject to limited transfers to sub-processors (source) | Stated by Mistral |
| SOC 2 Type II, ISO 27001 and 27701 | Mistral says it complies with them and provides its reports on request through its Trust Center (source). SOC 2 is an attestation report by an auditor, not a certificate; for ISO 27001 and 27701, we found no published certificate number or certification body | Stated, reports to request before signing |
| SecNumCloud | No published qualification for Mistral's own cloud. The qualification applies to a host: L'Assistant, the French State's tool, was built by the DINUM with Mistral AI on Outscale's SecNumCloud infrastructure (source) | A qualification of the host, not of Mistral |
| Training on your data | On by default on Vibe unless the user opts out; Enterprise opted out by default, setting managed by the admin; on Team, the admin can turn it off for the whole organisation (source, source) | Depends on your plan and settings |
| Thumbs up or down | Giving feedback with a comment allows Mistral to use the rating, the prompt and the answer, whatever the plan, Enterprise included (source) | A usage rule to write down |
| Models offered on the platform | Mistral also hosts open models from other vendors: Z.ai's GLM 5.3 and GLM 5.3 Flash, DeepSeek V4.1 Flash (source) | Model choice to govern |
When your teams put personal data into Vibe, Mistral acts as your organisation's processor. GDPR Article 28(3) then requires a contract setting out the subject matter, duration, nature and purpose of the processing, the types of data and the categories of data subjects: Mistral publishes a data processing addendum for this, to be reviewed with your data protection officer.
An assistant connected to email or to files about named people may create a high risk for individuals: Article 35 then requires an impact assessment before the processing. For transfers outside the EU, Article 46 allows, among other safeguards, the standard clauses adopted by the Commission, which Mistral says it attaches to its contracts with the sub-processors concerned (source).
For French State administrations, decree no. 2026-272 of 14 April 2026 governs particularly sensitive data entrusted to a cloud service provided by a private company. A French vendor does not amount to a qualification: for such data, the answer lies with a qualified host or with your own servers.
Six decisions shape what your teams will be able to do with Vibe. Each is checked against Mistral's documentation, reread on 30 October 2026, and settled with you during the briefing or the course.
In short, for leadership
Team is enough for a team whose data is neither sensitive nor about named people, provided the admin turns training off. Enterprise becomes necessary as soon as the organisation wants sign-in through its directory, audit logs, control over transfers outside the EU or a private deployment. For sensitive State data, the answer is a qualified host or your own servers; for files covered by professional secrecy, it lies in the settings, the contract and, depending on sensitivity, a model installed in-house.
The most important setting depends on the plan: check it before the first document, and separately for Vibe and for the API.
On Vibe, prompts and answers are used by default to train the models unless the user opts out; Enterprise customers are opted out by default, and only the admin can turn training on (source). According to the help centre article dated 3 September 2026, the admin of a Team plan can turn training off for the whole organisation, in the admin panel, under Vibe, then Privacy (source).
Documents uploaded to Vibe count as input data. Vibe and API settings are independent: turning one off does not turn off the other (source). If your teams already use ChatGPT with no framework, the ChatGPT training for companies covers the same move to a managed workspace.
Team lets you verify the email domain; sign-in through the company directory requires Enterprise.
Domain verification, through a DNS record, works from Team upwards; it allows anyone signing up with an address on the domain to join automatically, with the Member role, which Mistral recommends. Single sign-on via SAML, meaning sign-in with the company account through Okta, Microsoft Entra or Ping, requires the Enterprise plan (source). The Admin API, which lets you manage users, teams and permissions automatically, is also Enterprise only (source).
An agent only reaches what its connectors open to it: scope is set per workspace and tool by tool.
Since 24 June 2026, the admin sets connector access per workspace, for instance a finance workspace linked to internal sources with no web access, and turns each connector function on or off, whether to block all writes or a single delete action. API keys can be limited to one workspace's connectors. Mistral recommends that automated work run on behalf of a user or a service account, never impersonating the person who created it. A connector respects both the permissions of the source platform and those set in Vibe (source).
The catalogue lists more than a hundred connectors, including Outlook, SharePoint, Google Drive, Slack, Salesforce, HubSpot and ServiceNow; custom MCP connectors, meaning servers built for a business tool, are in beta (source, source).
Work asks the user's approval before sending, changing or deleting: decide which actions may go ahead without it.
Before any action that creates, changes, sends, posts or deletes in an external system, Work stops and offers three choices: continue for this action only, allow everything for the session, or decline. Each connector's functions are split into interactive and read-only ones; permanent approval is set function by function, and that setting belongs to each user. Mistral warns that the stop button does not guarantee interrupting an action already sent to the external system (source).
Skills carry the team's know-how, the Knowledge Base what Vibe remembers about each person: both need usage rules.
A workspace Skill can be used by all its members, and the admin can force-enable some of them so that users cannot turn them off. A Skill created in Mistral Studio can be published in Vibe (source).
The Knowledge Base is on by default. It records what the user asks it to remember and what it picks up by itself; sensitive personal information is stored only if the user turns on the matching setting. Turning it off does not delete existing entries, and server-level configuration is available for enterprises (source, source).
What Vibe does not yet allow: better to choose the plan knowing these limits than to discover them after purchase.
What we settle with you
Prepare these settings before the course
One first conversation is enough to agree the plan, the training setting and the migration timetable; the programme and quote follow within two business days.
Mistral publishes some of its models as open weights, meaning they can be downloaded and run freely: they can be installed on a workstation, a server or a private cloud. Here is the decision tree handed out in training, from the least sensitive data to the most protected.
| Your data | The right answer | To check |
|---|---|---|
| Public or internal, nothing about named people | Vibe Pro or Team | Training off on each account or by the admin |
| Personal data, across an organisation | Vibe Enterprise | SSO, audit logs, transfers outside the EU switched off, processing agreement and impact assessment |
| Covered by professional secrecy: lawyers, accountants, healthcare | Vibe Enterprise, configured and governed by contract, with anonymisation before upload; a model installed in-house for the most sensitive documents | Professional rules, vendor access set in the contract |
| Sensitive for the State, within decree no. 2026-272 | A SecNumCloud-qualified host, or Enterprise on-premises | Vendor access set in the contract: maintenance, support, data sent to it |
| Confidential, on one computer or a small server | An open model installed locally, for instance a Ministral 3 | Machine power, quality measured on your documents |
The Local AI training has a model installed and assessed on your machines; sovereign architecture consulting designs a full deployment; the AI for regulated professions training deals with professional secrecy. The Sovereign AI and local data guide explains how to choose a model for your hardware.
Beyond the migration steps shown above, this log lists five changes between 11 August and 29 October 2026; each was checked on 30 October in Mistral's own publications and chosen for what it changes in an organisation. The log is reviewed every month.
Mistral's largest model can be downloaded: free to use, modify and commercialise, except for a business reselling access to the model as a service. API price: $0.70 per million input tokens and $2.10 output. Mistral announcement
A one-trillion-parameter multimodal model, with 52 billion active parameters, trained in Mistral's European data centres. Mistral announcement
Z.ai's GLM 5.2 model, offered by Mistral since August, is deprecated and retired on 31 October in favour of GLM 5.3: an application calling it must switch models. Models page
Work generates dashboards and small tools in Canvas from a description; their data does not refresh by itself and must be regenerated. Release notes
Regional endpoints, Europe or the United States, are now open to all customers, and the platform starts offering other vendors' open models, starting with Z.ai's GLM. Mistral announcement
The course works through seven situations, each with the Vibe feature used, a sample prompt and the final check.
A workspace Skill holding the review checklist and the approved fallback clauses.
"Apply our review checklist to this supplier contract, list the clauses that depart from it with the relevant article and the planned fallback clause, and do not change the document."
Check: the lawyer reviews every deviation before any change to the contract.
Work on the month-end workbook, returning an Excel file.
"Reconcile the Bank tab with the General ledger tab in this workbook, list differences above €500 in a new tab with their likely explanation, and do not change any existing data."
Check: the controller approves each explanation before any entry is made.
A project folder with the job description and the collective agreement, and a Knowledge Base without sensitive data.
"Using the job description and the collective agreement in the project, write the job ad, then an interview grid with six observable criteria, and flag any wording that could be discriminatory."
Check: HR reviews the ad and the grid, and no candidate data is uploaded without a prior rule.
The meeting-prep Skill, linked to the calendar, email and CRM through read-only connectors.
"Prepare my Thursday meeting with this client from our latest exchanges and the CRM record: ten lines on the situation, the open points and three questions to ask."
Check: the salesperson checks the facts cited before the meeting.
One workspace per client, each with its brand guidelines written as a Skill, its documents and its connectors; the Knowledge Base, which is personal, is reviewed so that it does not mix information from two clients.
"Using the attached brief and this client's guidelines, suggest three campaign angles with a title and a hook each, adapt the angle I choose into a LinkedIn post and an email, then flag any departure from the guidelines."
Check: the project lead picks the angle and rewrites the chosen texts before anything is sent to the client.
A task scheduled every Monday and the internal-comms Skill.
"Every Monday at 8 a.m., gather last week's decisions and delays from the minutes in the Committee folder, and prepare a one-page note for the leadership committee, without sending it."
Check: the director reviews the note before it is circulated.
Vibe Code, in VS Code or in a remote session on the team repository.
"Explain the architecture of this repository's billing module, then propose a plan to add tests for the VAT calculation. Wait for my approval before changing any file."
Check: the developer reviews the plan, then each change before merging.
These prompts are starting points: in training, each is rewritten on a real document from your teams, then turned into a workspace Skill when the task recurs every week.
From €1,000 excl. VAT per half-day and €1,800 per day for an in-company group; final price depends on group size and length, on quote. Vibe subscriptions and API usage are paid by the client.
Colombani.ai is a Qualiopi-certified training provider (certificate QUA009472, training provider registration number 11757549975, this registration does not constitute State approval). Mistral programmes can therefore be submitted for OPCO funding, or FIF-PL funding for self-employed professionals, subject to the funder's decision. They are not eligible for CPF funding. For a public administration or local authority, training is paid from the employer's training budget; Colombani.ai provides the quote, programme and agreement.
Ulysse Trin
Engineer (Epitech, MSc AI), holder of the Claude Certified Architect Professional certification, issued by Anthropic, and of nine practitioner certifications issued by OpenAI, including ChatGPT Deployment Practitioner and ChatGPT Solutions Practitioner. Consultant, trainer and developer, he uses Claude and ChatGPT daily on client engagements and regularly tests Vibe, Copilot and Gemini to keep up with their changes; his programmes build on cases developed on those engagements.
Claude Certified ArchitectProfessional · Official Anthropic certification. Verify the credentialSee also the nine practitioner certifications issued by OpenAI → The coding agent comparison published by Colombani.ai → Training on Mistral AI, the French AI model vendor, and on its assistant Vibe, which replaced the Le Chat assistant. It comprises a half-day executive briefing, one day on Vibe Work for business teams and one day of local AI on open models. Colombani.ai is a Qualiopi-certified provider.
Not under that name: Le Chat became Vibe on 28 May 2026, and Le Chat Enterprise now corresponds to Vibe Enterprise. Chat mode is being folded into Work; for Enterprise customers, migration began on 1 October 2026, and the admin picks the date within a window of about six months. Skills replace agents and memories move into the Knowledge Base.
Mistral AI is a French company, and its customers' data is hosted in the European Union by default. Sovereignty, however, depends on the deployment mode: depending on the feature, data can travel to sub-processors outside the EU, which the Enterprise plan partly lets you switch off. For the most sensitive data, a SecNumCloud-qualified host or an open model installed in-house provides protection that the online subscription does not.
Mistral covers its transfers outside the EU with the safeguards of GDPR Article 46 and publishes a data processing addendum. Compliance nevertheless rests with your organisation: a processing contract under Article 28, an impact assessment under Article 35 where the processing carries a high risk, the training setting and a data usage rule.
As of 30 October 2026, Mistral lists, excluding taxes, Free at $0, Pro at $14.99 per month and Team at $24.99 per user per month; Enterprise is on quote. SAML SSO, audit logs and on-premises deployment exist only in Enterprise. List prices to recheck on Mistral's page before any purchase.
As of 30 October 2026, no SecNumCloud qualification is published for Mistral's own cloud. The qualification applies to a host: L'Assistant, the French State's tool, was built by the DINUM with Mistral AI on Outscale's SecNumCloud infrastructure.
On Vibe, prompts, answers and uploaded documents are used for training by default unless you opt out. According to Mistral's help centre, the admin of a Team plan can turn training off for the whole organisation, and Enterprise customers are opted out by default. Thumbs-up or thumbs-down feedback with a comment, however, allows use of the exchange whatever the plan. The course has these settings checked before the first document is uploaded.
From €1,000 excl. VAT per half-day and €1,800 per day for an in-company group; final price depends on group size and length, on quote. Vibe subscriptions are paid by the organisation.
An application can be made: Colombani.ai is Qualiopi-certified (certificate QUA009472). Funding is requested from the company's OPCO, or from the FIF-PL for self-employed professionals, and the decision rests with the funder. The programmes are not eligible for CPF.
Yes, in two ways: the Enterprise plan can be deployed on-premises or in a private cloud, and Mistral's open models can be installed on your machines. Locally, you lose the Work agent, its managed connectors and some quality, and maintenance becomes your responsibility. The Local AI day teaches you how to make that choice on your documents.
Both cover work on documents and code. For an organisation, the difference lies mainly in deployment: Claude is offered by Anthropic and on Amazon Bedrock, Google Vertex AI and Microsoft Foundry, with European regions, but not as an on-premises deployment; Mistral offers its assistant on-premises and publishes open models. Colombani.ai trains on both tools, and the executive briefing compares them on your cases before any purchase.
Not for the executive briefing, as the trainer runs the demonstrations. For the Vibe Work day, each participant uses a Pro, Team or Enterprise account and receives an installation guide three days before the session. For the Local AI day, no subscription is needed.
No. It is delivered by a provider that is Qualiopi-certified for training and leads to an end-of-training certificate listing the objectives achieved, without being registered with France Compétences. Colombani.ai holds no credential issued by Mistral AI.
Tailored programme and quote within two business days.
Describe your teams, the plan chosen or under consideration and your migration timetable: programme and quote within two business days, or a call at a time that suits you.
