On September 16, 2026, Mozilla announced Mistral Small 4 for Firefox Smart Window beta users in the US and Canada, alongside expansion to France with French-language support. Eligible Firefox users gain another assistant option inside the browser; this is an access and model-choice change, not a new launch of Mistral Small 4.
The practical question is whether help with pages you already have open is worth sharing their context. This guide checks access, model selection and data handling against official documentation available on October 6.
Where the beta is available
Mozilla’s getting-started guide describes a gradual desktop rollout requiring a Mozilla account: Firefox 150 onward in the US and Canada, and 155 onward in France. Meeting those requirements does not guarantee access.
The joint announcement named the UK and Germany as planned markets for later in 2026; the reviewed documentation does not confirm those launches. India is not among the confirmed markets. Check the official Smart Window page for the current access path.
Chat has daily usage limits that reset at midnight Eastern Time. Mozilla’s usage documentation does not specify a numeric allowance, and browsing continues after the chat quota is exhausted. The reviewed pages do not establish browser-specific paid pricing or future beta pricing.
Smart Window remains optional. Firefox’s AI Controls let you manage or block it; the partnership does not make Mistral compulsory for Firefox users.
Which model—and which host—are you choosing?
The Mistral-inclusive table in Mozilla’s model-choice documentation lists these options. The page also contains older model listings, so check your installation rather than assuming every rollout has an identical picker.
Assistant model | Model maker | Listed host |
|---|---|---|
Mistral Small 4 | Mistral AI | Mistral AI |
Gemini 3.1 Flash Lite | Google Cloud | |
Qwen3-235B | Alibaba | Google Cloud |
The in-chat selector changes that conversation only. To change the default, use Settings → AI Controls → Smart Window settings. Mozilla says its offered models follow the same data-handling standards and that providers do not retain conversations after processing. These commitments should not be assumed for a custom endpoint. Model selection and provider handling.
Model choice also has a feature boundary. Mozilla’s webpage-monitoring tasks documentation says those tasks use Qwen even when another assistant model is selected. That feature is documented for English locales in the US and Canada, with up to 15 active tasks and five pages per task. Selecting Mistral for chat does not make the entire browsing workflow Mistral-powered.
What the privacy commitments do—and do not—mean
Mistral and Mozilla say conversations are not saved on Mozilla’s servers by default and that partners including Mistral agree to zero retention. That is a published commitment, not an independent audit of processing or storage.
For Mozilla-provided models, the prompt, relevant memories and browsing context pass through Mozilla to the model provider. The Smart Window privacy notice also describes search-partner processing. Non-retention therefore does not mean that the content never leaves your device.
Mozilla says its proxy hides your IP address and browser/computer identifier from the AI service. Our interpretation: those protections limit identifying metadata, but a name or confidential detail included in a prompt remains content the model can process. Decide what may be shared before relying on a storage promise.
The sidebar automatically uses the current page as context; other tabs can be added with @ or +. Do not assume each page inclusion requires a separate confirmation. Chats are stored locally, so provider non-retention is not the same as having no chat history.
Two optional features need separate decisions:
Memories. Mozilla’s memory controls allow learning from chats and browsing in Classic or Smart Windows, excluding Private Windows. Mozilla processes inputs to create memories, then stores the memories on your device. Turning off memory use for one chat does not necessarily stop new memories being created. To stop learning, disable both browsing and chat learning; delete existing memories separately if you no longer want them used. Deleting history or chats does not delete memories.
Feedback. Submitted feedback can include the conversation, referenced page content and interaction details. Mozilla says an authorized team can review submissions and deletes them after one year. Preview the payload and exclude referenced pages when appropriate. This retention rule applies to submitted feedback, not ordinary chats.
For organizational use, separate the questions: what content is sent, who processes it, and what persists afterward? RohitAI’s guide to mapping AI data flows offers a broader review framework. Neither a European model supplier nor local memory storage, by itself, establishes EU-only processing.
When a custom endpoint helps
Mozilla documents custom remote and local models using an OpenAI-compatible API, including local servers such as Ollama and Lemonade Server. It warns that custom models may not work as expected. Compatibility is a starting requirement, not proof of equivalent browsing assistance.
The privacy notice says a custom endpoint outside Mozilla’s offered choices receives model requests without routing them through Mozilla’s server; a third-party endpoint’s own terms apply. A local chat endpoint therefore changes that inference path. It does not establish an entirely offline Smart Window: memory generation and search need separate review.
How to decide whether to try it
Try a narrow beta task if you regularly compare public pages and can access the feature. Mozilla documents sourced web answers and tab organization; those are concrete workflows to evaluate.
Keep Classic browsing with a separately scoped assistant if you want to hand-select every excerpt you share. The tradeoff to evaluate is manual context preparation versus control over the supplied material.
For confidential work, first establish permitted recipients and context. A preferred model name or a custom endpoint alone does not answer those questions.
Do not choose it for transaction automation. Mozilla’s current assistant guidance says it cannot click website buttons, fill forms or complete purchases. Page monitoring is a narrower capability.
A useful comparison, proposed here rather than performed: choose two public pages and one question with checkable answers. Give Smart Window and a standalone assistant the same source material, keep personalization off, and record context-preparation time, total task time, factual errors and unsupported claims. Check each answer against the pages. If changing models, repeat with fresh chats and the same inputs.
The adoption criterion is whether browser integration reduces the work needed to reach a correct answer under data-sharing rules you accept—not whether one model’s general benchmark score is higher.
Methodology: AI-assisted reporting and analysis based on the persisted research dossier and official announcements, support pages and privacy documentation rechecked on October 6, 2026. No hands-on browser testing, benchmark reproduction or independent privacy audit was performed.