Firefox adds visual history search and AI citations to Smart Window

Firefox released a major update to its Smart Window browsing mode today. The built-in chat can now pull up-to-date information from the web through a partnership with Exa and display source links directly in its answers. At the same time, the engine can suggest automatic tab groups, spot duplicate tabs and close them, and show visual previews of pages when users search their history with natural language—for example, asking “running shoes I checked last week” will surface images taken from the original sites.

The change shifts the AI’s role from generating fresh search results to sorting what users have already seen and adding citations from Exa. The announcement notes that this helps verify answers or return to a prior context when a user remembers a page’s content but not its URL. Beta testers reported that the feature helped them keep their train of thought, and future releases plan to add recent “browsing journeys” similar to Chrome and AI-driven autofill for online forms.

Smart Window remains an opt-in beta with no exact release date, according to Ajit Varma, head of Firefox at Mozilla. The feature, first announced last year as “AI Window,” now includes an AI Controls setting that lets users disable all AI capabilities or select which remain active. Varma described the approach as “agnostic,” offering tools to those who want them without imposing them on everyone. “They built a browser so you run more LLM queries; we’re looking at the browser experience itself,” he told The Verge.

Users can choose which model to run inside Smart Window: Google’s Gemini 3.1 Flash Lite, OpenAI’s oss-gpt-120b, Alibaba’s Qwen3-235B-A22B-Instruct-2507, or local models that run on the device. Varma said the goal is “digital sovereignty,” allowing selection of local or regional models with philosophies different from the major tech firms.

Steve Truong, senior product manager on Firefox’s AI team, emphasized that all offered models operate under zero-data-retention contracts: inference providers process prompts in memory only and do not store anything for training, publishing, or human review. Mozilla itself does not retain chats on its servers without explicit permission, nor does it use them for training, human review, or publication.

“We want it to be easy to get to the desired state, but not be prescriptive and say there’s only one way,” Varma concluded. Upcoming updates will continue this direction, adding more local organization tools and reducing the default push of generative chat.