Firefox’s new AI-powered ‘Smart Window’ feature is taking shape in development builds – I gave it a hands on in the latest v149.0b7 beta (on macOS).
Although Firefox 148 added an AI Switch, Mozilla is all-in on AI in now it’s part of its double-bottom line and a way to generate money. Firefox’s upcoming Smart Window is key to this, as it makes AI the primary interface for browsing.
What follows is a hands-on with Smart Window in Firefox 149 beta, how it works, how it’s activated and why its ‘memory’ feature gave me the ick.
Firefox’s new Smart Window feature is still under heavy development and is not yet enabled by default in beta builds. I couldn’t get it to run on Linux, but it did work on macOS. Nothing shown or mentioned in this article is finished or complete.
Firefox Smart Window
Smart Window is activated using a button in the window frame. Once activated, the browser’s colour scheme changes. I find it has echoes of the major Firefox Nova redesign that’s coming due to the roundness and pastel gradients.
Although presented as an entirely new mode for browsing, as opposed to an optional sidebar you might overlook, it currently retains the toolbar, address bar, extension main menu and other top-level items.
The first time you enter Smart Window mode, Firefox asks you to choose a model:
The beta doesn’t show model names on the selection screen, but they are they’re listed in Preferences > AI Controls > Smart Window:
- Fast – Google’s Gemini Flash Lite
- Flexible – Alibaba Cloud’s Qwen3-235B-A22B
- Personal – OpenAI’s GPT-OSS 120B
There’s also an option to bring your own LLM, with fields for model name, endpoint, and API token available for entry when the manual option is enabled. However, the page itself warns local models may not work correctly.
The new tab page drops the search bar and the “recommended stories” feed (ain’t no-one got time to read anything other than AI generated hearsay). In its place is a prompt box inviting you to “ask, search, or type a URL”, with suggested action chips beneath:
You can use prompt box like that of any LLM: ask it to help plan a meal, proof read text or generate homework festooned with LLM copywriting tics that “elevate” your writing using cliches like “it’s not X — it’s Y” and “whether you’re X, or Y…”.
Most pointed questions you type will start a Google search1. This loads a regular search results page, and sees Firefox’s AI chatbot shift to a sidebar on the right. The AI reads the top results (including any AI overview), and produces a response based on them.
You can click through to an actual web page, ask questions about it, and if the model decides it needs more information, it’ll run its own follow-up search using its own terms, in the same tab:
That’s the “smart” in “smart window”, presumably. It’s like an unsubtle version of the AI chatbot sidebar Firefox is forever nagging and calling attention to.
When using the prompt box (on the new tab page or in the sidebar) the action button becomes a menu with additional actions, depending on what you want to do:
Being self-absorbed, I threw Smart Window a few queries about this site (more because it’s easier for me to spot errors in summarisations and answers, given it’s my site and I’ve been writing it since 2009 – feels like only like yesterday I was unboxing a Dell Zino HD2).
The results were consistent in that responses were largely rephrased whatever was showing on the Google results page, padded out with inferences that sounded confident but often lacked substance (what we used to refer to as “BS”, but is now known as “you’re prompting it wrong”).
You can show/hide the AI sidebar at any time when viewing a regular web page, but it will re-open when you switch back to that tab. It might be a bug in this (early preview, did I mention?) build, or it might be a “pivot” annoyance.
What Smart Window can’t do is interact with websites directly. No form-filling, no clicking, no shopping. Given the legal trouble Perplexity’s Comet browser is in over ‘agentic’ behaviours, it’s an omission that feels the right call.
Managing memories (aka creepy recollections)
Smart Window uses ‘memories’, things Mozilla says “…it learns from your activity” to inform its responses.
You can delete memories individually, and you can set any given chat session to not use/store them.
Fine so far.
The problem? My memory list isn’t populated with things Smart Window learned since I enabled it. Oh no.
It has activity going back months. We’re talking searches and website interactions from long before I enabled this. features.
Firefox just handed that history to the AI models to plough from, without telling me upfront.
I found this the creepiest aspect of Smart Window.
Mozilla says this was a flub; it will refine the onboarding around Smart Window to limit memory formation to post-opt-in activity only. That’s obviously the right fix.
Because sharing a user’s prior browsing history with third-party AI models, silently, on feature activation, without any headset? Yeah, a bit icky – but that’s the price3 of testing features that are finished, I guess.
Who’s hosting the LLM models?
Off the back of learning my entire search history for the past year or so had been handed over to an AI model, I thought to myself: “They know me intimately, why don’t I get to know them?”.
So I asked each model where it (thinks) it is hosted/running.
Gemini (Fast) could not be coaxed, and would only deflect to say:
I am a large language model, trained by Google.
Qwen (Flexible) was more forthcoming, emphasis mine:
While I am integrated into Firefox as Smart Window by Mozilla, the underlying Qwen model that powers me runs on infrastructure provided by Alibaba Cloud. This means my AI capabilities are hosted and operated by Alibaba Cloud, while Mozilla provides the browser integration and user interface within Firefox.
GPT-OSS (Personal) was also upfront:
This model runs on OpenAI’s cloud infrastructure. Mozilla integrates the service into Firefox, but the inference (the actual generation of responses) happens on OpenAI’s servers, and the results are then passed back to the browser for you to see.
Can we trust those responses, or is it more mirage with plausible-sounding responses?
Still, the answers are consistent with the obvious fact that Mozilla can’t afford4 to run major LLMs on its own infrastructure at scale, so is mediating access to Alibaba, OpenAI, et al.
That could have implications on where your questions (and potentially your browsing history) end up, and which governments or agencies have access or oversight of it.
Keeping an eye on being kept an eye on
I tested the feature on Firefox 149.0b7 on macOS as enabling the various Smart Window feature flags didn’t work (for me) on Linux. I got the window switcher button to appear, but the feature itself wouldn’t do much.
So if you’re on Linux and curious, it’s one to watch.
Do I think this feature is going to undo years of decline in Firefox’s marketshare? No. It’s another case of copying someone else’s homework and making it look a bit different (much like generative AI in general).
Would I use this feature personally? Again no. AI-as-default-browsing-interface is not something I want or have use for.
If I wanted someone else to read the web on my behalf, parse what it says on topics I care about and then confidently BS to me about what’s happening, I’d go and watch YouTuber.
But at least now you have a better idea of what’s coming.
- One major danger in cauterising the web of human content is it gives more weight to the super-funded. Bad-actors can “flood the zone” – if enough websites mention something, an LLM is inclined to assume it’s probable, and is happy to regurgitate it with confidence. ↩︎
- An hallucination on my part. ↩︎
- Now Alibaba knows I love Beavis and Butthead. ↩︎
- Mozilla may be making bumper revenue since it laid off staff, axed common good projects and launched an ad-tech arm, but it’s not making ‘AI gold rush’ money. ↩︎