

Mozilla and Mistral put model choice inside Firefox. Privacy still needs reading
Mozilla has made Mistral Small 4 the personalised option in Firefox Smart Window. That is a meaningful move for model choice, but it is not a promise that browsing data disappears.
By Katie Delaney / 2026-09-16 / 10 min read

A Firefox AI assistant with another door
Mozilla and Mistral have announced a partnership that places Mistral Small 4 inside Firefox Smart Window beta. The feature is not a replacement for Firefox as most people use it. Smart Window is an optional, separate browsing experience, and Mozilla says Mistral joins other selectable models rather than becoming a compulsory browser default. Mozilla’s partnership announcement
The practical change is regional and gradual. Mozilla says Mistral Small 4 is available to Smart Window users in the United States and Canada, while France gains Smart Window beta access with official French-language support. The partnership page describes further European expansion as planned, not delivered. That distinction matters more than the rhetoric about an open AI future.
Firefox’s own support material describes Smart Window as a Firefox AI assistant with three hosted model options and an advanced route for bringing a custom model. A user can change the default in AI Controls or choose a different model for one conversation. Firefox model choices documentation That is a more concrete form of choice than a browser sidebar that merely links out to competing chatbots.
Mozilla selected Mistral Small 4 after its own Smart Window evaluation, including multilingual performance. That is a vendor selection, not a public benchmark result. Still, it explains why this partnership is principally a distribution story: Mistral reaches a browser audience without owning the browser, while Firefox can recommend a European model without forcing every user into it.
Hosted choices
Firefox support documentation
Initial launch regions
United States and Canada
First new market
France, with French-language support
Choice is not a privacy policy
The strongest version of this story is not that Firefox has solved browser AI privacy. It is that Mozilla has put a model selector next to a browser feature that can see useful context. That makes the privacy question more legible, but it does not make it disappear.
Mozilla’s Smart Window privacy notice says the feature is opt-in. It also says that, when enabled, it collects and processes information about browsing activity to provide AI-enhanced features. The assistant can use a query, relevant Memories and browsing context. Mozilla Smart Window Privacy Notice That is more precise than describing the product as private by default.
There is a meaningful safeguard in the architecture. Mozilla says the full prompt is sent to a Mozilla server, which forwards the request to the selected large language model. The provider sees a Mozilla IP address rather than the person’s own IP address, and Mozilla says the request contains only the information needed to answer. But a prompt can still contain sensitive context if a person has permitted that context to be used.
Mozilla says model makers do not store conversations after brief processing. It separately says Memories are stored locally on the device and expire unless later activity keeps them relevant. Those are useful boundaries, not a reason to skip the settings. Users can review, delete or exclude Memories from a chat, and can return to a classic Firefox window at any time.
Opt in to Smart Window and select a hosted or custom model.
Allow, remove or disable Memories and choose which browsing context is useful.
Mozilla forwards the necessary prompt to the selected model provider.
Inspect cited sources, Memories and settings before treating an answer as settled.

Why Mistral fits the Firefox frame
Mistral Small 4 gives Mozilla a model story that is about more than speed. Mistral describes it as multimodal, multilingual and available under Apache 2.0, with a 256,000-token context window and configurable reasoning effort. Mistral Small 4 technical announcement Those are provider specifications, not an independent scorecard, but they explain the partnership’s emphasis on language and adaptability.
Mozilla’s support documentation is more revealing than a launch slogan. Firefox says it evaluates models on common browser tasks, trust and safety, usefulness in real browsing scenarios and usage limits. It also distinguishes a model’s promoted characteristic from a universal ranking. Firefox’s evaluation criteria
That makes Mistral Small 4 a recommendation, not a monopoly. A person who prefers Google Gemini, Qwen or a custom endpoint retains a route to it. The choice is particularly important for a browser AI assistant because browsers sit beside work, shopping, research and private reading. A model picker cannot neutralise every data risk, but it can stop one company’s default from becoming an invisible gatekeeper.
The terms need reading
Some of the unease around browser AI is less about Mistral than about the terms and notices that accompany an assistant able to work across tabs, history and remembered preferences. That unease is rational. A product can be optional and still deserve close scrutiny once it begins to infer interests from browsing behaviour.
Mozilla’s notice says it may infer location-related information when a query appears to contain it, and may use page titles, URLs and relevant page content to help answer a request. Private Windows are excluded from Memories. The notice also says that a custom endpoint is not routed through Mozilla and is governed by that endpoint’s own terms and privacy notice. privacy controls and custom endpoints
That is the article’s central caveat. ‘Zero retention’ is too blunt. Mozilla says it does not retain chat logs without permission and that model providers do not keep conversations, yet Smart Window processes context and stores Memories locally. The more useful claim is narrower: the product offers controls, a proxy layer and local memory management, while asking people to make informed choices about a context-aware assistant.
For a virtual assistant, consent cannot be a one-time ceremonial click. It needs a working interface: visible model choice, an understandable explanation of what context is used, a way to remove a memory and a clear route back to an ordinary browser window. Firefox has placed those levers nearer to the feature than many browser integrations. The test is whether users can actually understand and use them.
| Claim | What the official notice says | Useful user action |
|---|---|---|
| Conversations are not retained | Model providers briefly process but do not store them | Check the selected model and provider |
| Memories are local | Mozilla processes data to infer interests, then stores Memories on device | Review, delete or block Memories |
| Custom endpoint means control | Data goes to the chosen endpoint and its terms apply | Read that endpoint’s privacy notice |
- Conversations are not retainedModel providers briefly process but do not store themCheck the selected model and provider
- Memories are localMozilla processes data to infer interests, then stores Memories on deviceReview, delete or block Memories
- Custom endpoint means controlData goes to the chosen endpoint and its terms applyRead that endpoint’s privacy notice

What to try before you trust it
A Firefox AI assistant is worth trying as a bounded experiment, not a migration of every browsing habit. Start with a low-stakes task: a set of public tabs, a comparison you can independently check, or a page you want summarised. Watch which context it says it used and whether the cited sources hold up.
Then change one variable at a time. Try Mistral Small 4 for a multilingual task. Try another hosted model for the same question. Turn Memories off and repeat. Use a classic window for anything that should not be part of a context-aware project. This is not paranoia. It is how people learn what an assistant actually does with browser context.
For teams, the commercial lesson is equally plain. Browser AI is becoming a distribution layer for model choice. That makes clear sources, structured content and accountable privacy explanations more valuable. A brand may be discovered through a browser assistant long before a user visits a conventional search page. folkfox’s AI visibility dashboard guide is a useful starting point for measuring that shift without treating it as magic.
The partnership is a meaningful signal for competition, particularly because Mistral is a European model company and Firefox remains independent. But the success measure is not whether Smart Window sounds progressive. It is whether people can choose a model, understand the data route, inspect the answer and leave when the feature is not right for the task.
The keyword evidence is deliberately restrained. DataForSEO returned one close, measurable discovery term, ‘virtual assistant’, at 5,400 monthly searches and difficulty 27, but did not return a usable low-difficulty cluster specific to Smart Window. This Firefox AI assistant article therefore prioritises exact product language over stuffing unrelated browser terms into the copy.
The documentation deserves a slower read than the launch cycle permits. Firefox’s safety guide describes a Mozilla proxy intended to shield direct browser identifiers and IP addresses from an AI service. Its Memories guide says that browsing and chat history can be briefly processed to make local summaries, while the custom-model guide explains that a compatible endpoint may be local or remote. Those distinctions are the scent trail a careful reader should follow. A Firefox AI assistant may reduce one kind of exposure while creating a different decision about context, account use or a third-party endpoint.
There are other practical boundaries. Mozilla’s feedback guide says submitted feedback can include reviewed chat material, while ordinary chats remain separate. The Smart Window terms sit alongside the feature-specific notice, and Firefox’s wider privacy notice distinguishes on-device AI features from services that integrate third-party models. None of that makes the product a trap. It makes the setting screen part of the product. The sensible approach is to keep the work small, track the context you have allowed, and use the ordinary browser whenever a task does not need an assistant. That is not a burrow of fear. It is basic control.
For a Firefox AI assistant, the strongest promise is reversible use. A person should be able to turn it on for one project, inspect the answer, remove the Memories that no longer help, and switch back without losing their bearings. Mozilla’s AI-controls overview presents exactly that model of selective use. Whether the execution matches the intent will be tested in the beta, by people following the trail from a useful answer to its context, its sources and its settings. That is where trust is earned, one cautious paw at a time.
The practical test for a Firefox AI assistant is repeatable. Use the Firefox AI assistant on a public research question, then run the same task in a classic window. Use the Firefox AI assistant with Memories disabled, then compare the answer to a new chat with a different hosted model. This does not turn every reader into an auditor. It asks the right ordinary questions: did the Firefox AI assistant cite a source, did it use only the context expected, and is the answer still useful when the remembered material is removed? A Firefox AI assistant becomes easier to trust when those answers are visible rather than assumed.
The important browser AI feature is not the model selector alone. It is the ability to see what the assistant can use, change it and walk away.
Frequently asked questions
What is Firefox Smart Window?
Smart Window is Firefox’s optional, beta browsing experience with a context-aware assistant. Mozilla says it can work with open tabs, browsing history and Memories when a person permits that use.
Which model does Firefox recommend after the Mozilla and Mistral partnership?
Mozilla says Mistral Small 4 is the personalised model in Smart Window in supported launch regions. Users can change the model in Settings or for an individual chat.
Does Smart Window store conversations?
Mozilla says model providers briefly process a conversation to answer, but do not store it. Mozilla says it does not retain chat logs without permission. This is separate from local Memories and browser context used to tailor a prompt.
Can I use another model in Smart Window?
Yes. Firefox documents hosted choices and an advanced custom-model route. A custom endpoint processes information under its own privacy terms and does not route through Mozilla.
Is Smart Window available everywhere?
No. Mozilla describes a phased beta rollout. The partnership announcement names the United States, Canada and France, with additional European expansion planned rather than confirmed.
How can I reduce the context used by the assistant?
Review or delete Memories, block them for a chat, close relevant tabs, remove pages from history or bookmarks, and use a classic Firefox window. The exact choices depend on the task and feature settings.
The Agents API makes autonomy a governance decision
Read the field notefolkfoxHow to build a meaningful AI search visibility dashboard
Read the field notefolkfoxGenerative engine optimisation services: what AI Overviews can actually cite
Read the field notefolkfoxOpenAI for Finance: the sharp test for AI agents
Read the field noteNeed an AI claim that can survive a close read?
folkfox turns fast-moving technical evidence into clear strategy, useful editorial and accountable decisions.
Want folkfox in your Google results and AI answers? Set folkfox as a preferred source.

