Mistral trains on your chats by default. Turning it off takes three separate switches

Mistral's data-governance pages carry an "Updated this week" stamp as of 2 September 2026, and one line on them is worth reading twice if you use the company's assistant without an enterprise contract. On Vibe, the product that until recently was called Le Chat, "input and output data are used by default to train our artificial intelligence models, unless you opt out of training at any time."

The enterprise tier gets the opposite default, and Mistral's wording there is just as plain: "Enterprise customers are opted out of training by default. The opt-in toggle is managed at the admin level."

Both sentences sit on the same page, one under the other. The protection follows the contract rather than the sensitivity of what you typed.

The product changed its name, and that is where the setting lives#

Le Chat is now Vibe, which Mistral describes as a unified agent for professional productivity and coding. It comes in three modes: Vibe Work for delegating multi-step tasks across your apps, Vibe Code as a command-line tool and VS Code extension, and Vibe Chat for the turn-based conversation Le Chat was known for. The URL and the login are unchanged, and Mistral says accounts, plans, conversations and saved content all moved across.

That matters here for a practical reason rather than a branding one. The training toggle is filed under Vibe in the admin panel, and the detailed product documentation has moved to docs.mistral.ai, so searching for the old name will land you in the wrong place.

There is no single switch#

Mistral publishes three separate opt-out paths. You need whichever ones match how you actually use the product.

Vibe on the web. Open the Admin panel, select Vibe under the Manage section, then under Privacy disable the toggle labelled "Allow your interactions to be used to train our models".

Vibe on iOS and Android. Open the app's Settings page, select Data & Account Controls under the Account section, then deselect the "Enable data sharing" checkbox.

Mistral Studio and the API. In the Admin panel, open the Privacy menu in the left-hand navigation bar, then under "Anonymous improvement data" disable the toggle that allows API calls and related data to be used to improve Mistral's services.

Then comes the sentence that turns this from a note into an article. Mistral states it directly: "Vibe and API opt-out toggles are separate. Opting out of one does not opt you out of the other". Each toggle, it says, has to be configured individually.

That is the trap, and it is an easy one to walk into. A developer who switched the API toggle off in Studio, reasonably assuming it covered the account, is still contributing training data every time they open the chat interface with the same login.

Uploaded files count as well. Mistral spells that out: "Documents attached or uploaded within Vibe are considered as input data." The opt-out is not only about what you type, it also covers the contract you dropped into the window to be summarised.

The thumbs button is a separate authorisation#

This is the part with teeth, and it is not a setting.

Feedback sits outside the opt-out entirely. Mistral's page says that "in any case, and whichever subscription plan you adhere to", using the thumbs up or thumbs down button and adding a comment authorises the company to use your rating, your input and your output to improve how its models behave.

Read the scope of that clause: whichever plan. Enterprise included, opted out or not. Mistral's own advice is the clearest statement of the consequence, and notice that it is advice rather than a control: "Please refrain from providing feedback if you do not want us to use your feedback data."

So the rating widget is not a quality signal that happens to get logged. It is a consent action, and no toggle undoes it. If you have opted out, the single habit worth breaking is the reflex thumbs up on a good answer.

Zero Data Retention is a different control, and probably not yours#

Anyone reaching for Zero Data Retention as the stronger version of an opt-out should know that Mistral treats the two as unrelated. Its page draws the line explicitly: ZDR governs whether supported API inputs and outputs are retained after processing, while the training opt-out governs whether eligible data may be used to improve the models. Switching on one tells you nothing about the other.

Four limits sit on it, all of them stated by Mistral:

• It is available only with pay-as-you-go. On Free mode you have to enable pay-as-you-go first.
• It covers stateless API endpoints only, and Mistral lists them: chat completions, FIM completions, embeddings, moderations and classifications, OCR, speech and transcriptions.
• It does not apply to stateful products at all. Mistral names agents, batch processing files, conversations, libraries, the Files API, Vibe Work and Chat. Uploaded files are outside its scope by design, because the Files API stores them.
• It is not a switch you flip. You submit a request giving "sufficient detail of your legitimate reasons", and Mistral reviews it and approves or denies at its discretion. An approved request appears in the Privacy settings of the Admin panel, and a rejected one arrives by email.

For Vibe there is no ZDR at any price. Mistral's reason is architectural rather than commercial: it has to store conversations to show you your history and keep application state. Conversations are removed on manual deletion or when you delete your account, after what it calls a short grace period.

Team and Enterprise administrators do get company-wide controls, though the ones Mistral lists are about access rather than training: single sign-on, user management with permissions, and an organisation-level opt-out of its Flash Answers feature.

What to actually do#

If you use Vibe on a personal or Pro account, flip both toggles today, the web one and the mobile one, then stop using the thumbs buttons. That closes the default that is actually costing you something, it takes about a minute, and nothing about the product gets worse afterwards.

If your team is on Studio or the API, treat the toggles as two jobs rather than one, and check the Vibe side even when nobody believes they use it, because the same login reaches both. The wider version of that exposure, and how to price it before it becomes a procurement problem, is the subject of the business risk in open versus closed models.

And if the material genuinely cannot leave your control, a vendor toggle is the wrong instrument for the job. A setting is a promise about behaviour that a policy update can revoke; running the model yourself changes what is possible rather than what is permitted. Local versus cloud processing sets out that trade honestly, including the parts where local loses, and the security risks of running a local LLM is the counterweight to read before assuming local means private.

None of this is a scandal. Mistral documents its defaults in public, in readable language, with the enterprise asymmetry stated rather than buried. What your AI coding assistant sends home asks the same question of the tools most developers leave running all day. The point here is narrower and more useful than outrage: the default is on, the switches are plural, and the one thing that overrides all of them is a button designed to feel helpful.