Firefox AI Smart Window Adds Mistral With Zero Data Retention

What happened

Mozilla's experiment with AI features that people can turn down came together this autumn, and the pieces now fit. On September 16, Mozilla and the French AI company Mistral announced a partnership that brings Mistral's models into the Smart Window beta, Firefox's AI browsing assistant. The assistant can work across open tabs and browsing history to help with complicated searches and pull information together.11 Both companies say the arrangement uses zero data retention. Smart Window conversations are not stored on Mozilla's servers by default, and Mistral has committed not to keep user data for training its models.15 Two weeks later Mozilla shipped a redesigned Firefox. The redesign brought back Compact Mode, made privacy settings easier to find, and added a single control that turns off every generative AI tool.19

None of this is a breakthrough in model capability. What sets it apart is the product approach. Firefox now offers an AI assistant that can be switched off completely, that is honest about where data goes, and that runs on an open-weight model from a smaller European vendor. The rest of the industry is moving toward autonomous agents that critics have shown to be fragile. Whether that is enough to help a browser with a shrinking market share is a different question, and the coverage disagrees on it.

Inside Smart Window

The model behind the deal is Mistral Small 4, released under the Apache 2.0 license and available on Hugging Face. Mozilla reportedly picked it for its multilingual abilities.15 Mozilla treats Smart Window as a third kind of window, alongside regular and private windows. It combines tab analysis and automatic grouping, an AI chat, and AI answers backed by web search.15 One account of the beta adds that it uses the Exa search service to return current answers with links. It also removes duplicate tabs, shows a visual preview of history, and lets users manage a locally stored assistant "memory."20

Accounts of the rollout differ a little. One report says Smart Window is rolling out in France and North America.11 Another says the Mistral model now powers the beta in the US and Canada, while access and French-language support expand in France.16 The UK and Germany are expected later this year.12 A guide based on Mozilla's support documentation lists Firefox 155 as the version needed in the US, Canada, and France.17 The scale is modest. One estimate says the deal could add about 200,000 monthly active Firefox users to Mistral's reach.16

One detail most coverage skips matters for the privacy claims. Smart Window does not use only Mistral. Firefox documentation lists Gemini 3.1 Flash Lite and Qwen3-235B as other selectable models, which users can change under Settings > AI Controls > Smart Window settings.17 Mistral describes its zero-retention commitment as applying to the Smart Window partnership in France and North America.17 Mozilla's broader promises are that it does not keep chat transcripts, does not train models on conversations, does not sell user data, and shares chat content only when users choose to.16 The sensible conclusion is that the strongest guarantee applies to the Mistral option specifically. Users who care about the details should check which model is selected.

The off switch came first

The Mistral announcement only makes sense alongside what Mozilla did earlier in the year. In December 2025, Mozilla's new chief executive, Anthony Enzor-DeMeo, set a more AI-focused direction for the company and promised an AI "kill switch" to users who were unhappy about it.28 He argued that showing a commitment to choice is how a company earns trust.28 At the time, skeptics pointed out that the promise was a policy statement with no shipped code behind it. They also noted that Mozilla had not said whether disabled AI components would be fully removed or just hidden from the interface.30

Firefox 148 delivered the switch in February. A new AI Controls panel offers a "Block AI Enhancements" option that turns off AI features all at once, or users can toggle them one at a time.23 The controlled features at launch were translations, alt text for images in PDFs, AI-assisted tab grouping, link previews, and the chatbot sidebar.24 Mozilla committed that these preferences would survive updates, so the browser would not quietly turn AI back on.22 Privacy Guides pointed to a quieter improvement: the panel labels which features run on the device and which send data elsewhere, such as the chat sidebar.24

The local features have a reasonable privacy record. Translations run on the device instead of in the cloud.21 The PDF alt-text generator also runs locally, and Mozilla says the data is not used to train models.21 Tab-grouping suggestions run on a local model on the user's computer.25

The coverage disagrees on what the defaults were. Privacy Guides wrote that AI features are blocked by default and must be opted into.24 How-To Geek, writing before the release, said these features had usually been enabled by default and had to be switched off one by one as they arrived. It also recalled an AI-powered history search that briefly caused battery drain.25 The best explanation is that the kill switch partly fixed earlier missteps. That makes the switch more believable, because Mozilla appears to have reacted to complaints from its own users. It also explains why longtime users are still wary.

The switch has limits. Mozilla's help pages say the AI Controls panel does not cover features that use traditional machine learning to classify, rank, or personalize.26 "Block AI" therefore means blocking generative AI, not every model in the browser.

Why this matters: the agentic alternative

The comparison with AI-first browsers is what makes Firefox's approach stand out. Perplexity's Comet and OpenAI's ChatGPT Atlas were built around agents that read pages and then act on them, filling in forms or going through checkouts with little supervision.2 Security researchers have spent a year showing that this is where the weakness lies. Brave demonstrated that hidden text on a webpage, including invisible instructions inside a Reddit spoiler tag, could make Comet extract a user's email address and one-time passcode.5 Within hours of Atlas launching in October 2025, researchers showed that content in a shared Google Doc could steer it.2 OpenAI has said that prompt injection is unlikely ever to be fully solved.5 University of Washington researchers found that four of seven agentic browsers they tested allowed attackers to bypass the same-origin policy.7 In one phishing test, Comet blocked 7 percent of attacks and Atlas 5.8 percent, compared with 53 percent for Edge and 47 percent for Chrome.6

One report says OpenAI planned to retire Atlas on August 9, 2026, with security concerns among the reasons, and to move its agent features into other products.3 Even if that holds, the risk continues wherever agents carry over users' logged-in sessions.3

Smart Window works differently. As described, it organizes, summarizes, and helps users find things, and it does not take actions on users' behalf.15 Privacy Guides has urged Mozilla to keep to local, non-agentic features. It warned that agentic AI is the next trend and that Mozilla should be careful.24 So far, Mozilla's main AI product has followed that advice. A separate 2026 browser security review gave Firefox credit for unusually clear, centralized AI controls, while noting that Chrome and Edge offer more powerful agentic capabilities that need more careful permission management.14

The smartphone angle

The same idea has reached mobile. Firefox's Android listing now advertises "AI on your terms," which lets users turn AI off entirely or adjust it feature by feature. The listing sits next to a shake-to-summarize gesture that pulls out a long page's main points.13 Phones are becoming the main place where AI assistant apps compete, so a browser that offers a summarizer without making it mandatory is a small but noticeable contrast.

The skeptics' case

The coverage is not all positive. A long critique on HackerNoon argues that Mozilla's AI push repeats an old habit of chasing trends instead of building a clear identity. The same piece concedes that local tab AI and a choice of providers are improvements, and that the controls Mozilla added in February and reinforced in September present AI as optional.18 One analysis of the Mistral deal notes that Firefox's market share has been falling steadily, and says it remains to be seen whether AI can turn that around.15 Privacy Guides also criticized Mozilla's investments in more than 50 AI companies, a so-called "rebel alliance," for offering little technical detail on privacy.24 On Slashdot, readers argued over whether the switch was a real safeguard or a fix added after the fact.26

The verdict

The complaints are valid, but they do not change the main point. Firefox has shown that a mainstream browser can include AI features while letting users disable all of them, label what stays on the device, keep that choice through updates, and pair its flagship assistant with a vendor that has publicly committed to zero data retention. One commentator called the redesign unusually clear-headed in a market tired of AI.19 TechRadar hoped other developers would follow.27 Mozilla's goodwill now rests on keeping that switch reliable as Smart Window grows and agentic features become more tempting.