Configuring Firefox Smart Window: Models, Memory, and Privacy Controls

To test Firefox Smart Window, it is essential to install the beta version of the browser, specifically 149.0b7, where the feature has been publicly released for the first time in a stable manner on macOS. Current availability is limited: the experience works with relative stability on macOS, while on Linux it remains erratic and is not considered suitable for daily use. In September 2026, Mozilla expanded the beta to the United States, Canada, and France, incorporating official French support for the latter market.

Activation is performed via a button integrated into the browser window frame. When pressed, the interface changes completely: a new color scheme is applied that anticipates the Nova redesign of Firefox, and the new window tab transforms. Instead of the typical page with a search bar and shortcuts, a single central field appears inviting you to ask, search, or write a URL. This design unifies classic search, AI chat, and the address bar into a single entry point, simplifying initial interaction and reducing visual friction for the user.

Mozilla does not commit to a single artificial intelligence provider. In the initial tests of beta 149.0b7, the user can choose from three preconfigured options, each with a different focus on latency and privacy. Additionally, there is the possibility to connect a custom model via a custom endpoint, which opens the door to enterprise deployments or self-hosted models, offering considerable technical flexibility.

The choice of model depends on the balance you seek between speed, versatility, and customization. The main options documented in the beta are summarized below:

Preconfigured Options

• Fast (Gemini Flash Lite): Based on the Google model, it prioritizes low latency for quick responses. It is ideal for simple queries where immediacy is the critical factor.

• Flexible (Qwen3): Developed by Alibaba Cloud, it offers a balance between reasoning capability and adaptability. It is positioned as an intermediate option for tasks requiring certain logical complexity.

• Personal (GPT-OSS): Based on the OpenAI model, it focuses on interaction personalization. It seeks to adapt the tone and structure of responses to user preferences.

• Mistral Small 4: Incorporated in the September 2026 expansion for the US, Canada, and France, it adds an open AI alternative. This inclusion reinforces Mozilla's bet on provider diversity and access to open-source models.

One of the most sensitive aspects of Smart Window is its memory system, designed so that the assistant remembers part of your activity and offers more personalized responses. However, in the initial tests, a critical problem emerged: the memory list included activity from months prior to the explicit activation of the feature. This meant that the system was processing information from the history before having the user's conscious consent, generating doubts about privacy.

After the situation became public, Mozilla acknowledged the flaw and committed to adjusting the activation process so that memory only begins to form from the exact moment the user accepts the feature. To maintain control, it is fundamental to periodically review which memories the assistant stores, delete those you do not wish to keep, and, if necessary, turn off the memory function entirely. Transparency in consent and the option to delete are key to mitigating community distrust and guaranteeing an experience respectful of personal data.

From Firefox 148 on desktop, the browser incorporates a specific section of AI controls within the settings. In that panel, a key switch appears: Block AI enhancements. When activated, Firefox disables all AI-related functions at once, including pop-ups, reminders, and suggestions. It is, in essence, a panic button for those who do not want automations and prefer a traditional browser without algorithmic interference.

In addition to the global block, the panel allows managing granularly which AI-based tools you want to have active. You can disable automatic translation, alternative text generation in PDFs, intelligent tab grouping, or smart link previews. Preferences are respected between updates, so functions are not reactivated without your explicit authorization. This architecture allows activating only what is useful and leaving the rest off without complications, granting the user fine control over their digital experience.

It is important to understand that Smart Window does not go as far as navigating for you or acting as an autonomous agent. It cannot perform actions on other websites on your behalf: it does not fill out forms, does not click buttons, nor does it automate complex processes. It is limited to reading what is in the open tabs and using that information to respond, which greatly reduces the legal and security problems faced by more agentic browsers and their ethical implications.

In practice, the AI's behavior tends to paraphrase what already appears in search results rather than offering revolutionary ideas. It is useful for having a distilled and well-ordered view of a page, but it does not replace direct reading if nuances or fine details are sought. The feature is designed as a trusted companion, not as a substitute for the open web, and its objective is to help discover and understand content, not to enclose the user in a conversation loop.