Can Firefox Smart Window Run Fully Local? What Ollama Changes—and What Still Leaves Your Device

Short answer: Firefox Smart Window can point its built-in assistant at a local OpenAI-compatible model endpoint, including Ollama or Lemonade Server. That can keep the assistant model inference on the same device.

But “I selected Ollama” is not the same thing as “Smart Window never sends anything to the internet.” Smart Window is a bundle of separate features. Mozilla documents different data paths for assistant chats, web search, memory generation, browser history, and local storage.

That distinction matters because a privacy-conscious setup should be configured by data path, not by one model dropdown.

Feature check — August 23, 2026: Smart Window is still an early beta for Firefox desktop. Mozilla says rollout began gradually for users in the United States and Canada with Firefox 150, with France starting in Firefox 155. Smart Window requires a Mozilla account. Custom models can be remote or local, but Mozilla warns that custom-model behavior may not match the built-in experience exactly.

The useful mental model: Smart Window has four separate data paths

Instead of asking “Is Smart Window local?”, split the feature into four questions:

Where does the assistant model run?
Where are memories generated?
What happens when the assistant searches the web?
Where are chats and memories stored afterward?

Those answers are not identical.

Smart Window function	Built-in setup	Local-model setup	What to know
Assistant model inference	Requests are routed through Mozilla before reaching an AI provider	Can be pointed at a localhost OpenAI-compatible endpoint such as Ollama	Local inference changes the model path, not necessarily every other Smart Window service
Web search	Uses Mozilla's Exa integration when current external information is needed	The browser may still use the Smart Window web-search feature	Local LLM does not automatically make web search local
Memory generation	Mozilla servers briefly process selected browsing/chat activity, then return memories to the device	This is a separate Smart Window feature	Turn learning off if server-side memory processing is not wanted
Chat and memory storage	Stored locally on the device	Stored locally on the device	Local storage can still be visible to other people who share the device
Private-window activity	Excluded	Excluded	Mozilla says Smart Window does not use Private Window activity for memories

The practical conclusion is simple: local-model mode is one privacy control, not a master offline switch.

What does choosing Ollama actually change?

Mozilla's current custom-model documentation lets Smart Window connect to any server that exposes an OpenAI API-compatible endpoint.

For a local Ollama setup, Mozilla's example uses a localhost endpoint similar to:

http://localhost:11434/v1

Firefox sends assistant requests to that endpoint instead of using one of Smart Window's built-in model choices.

The same idea works with Lemonade Server. Mozilla currently documents a localhost endpoint such as:

http://localhost:13305/api/v1

For the assistant-model step, this can keep inference on the computer running Firefox.

That is meaningful. A page summary, comparison, or question that would otherwise need a remote model can be handled by a model hosted locally.

But it is important not to overstate the result.

Mozilla's documentation treats custom model choice, memories, and web search as separate Smart Window features. Choosing a localhost model does not, by itself, say “disable Exa,” “disable memory generation,” or “turn Firefox into an offline browser.”

What still may leave the device?

1. Memory generation

Smart Window can create short personalized memories from two optional sources:

browsing history from Classic and Smart Windows;
Smart Window chats.

Mozilla says the initial memory process may consider up to 60 days of browsing activity or 3,000 items, whichever limit is reached first.

The resulting memories are stored locally. However, Mozilla says the information used to generate them is briefly sent to Mozilla servers for processing. Mozilla says it does not retain or log that activity beyond what is needed to complete the request.

So a user who wants local assistant inference but does not want browsing/chat activity processed by Mozilla for memory creation should also disable:

Learn from chats in Smart Window;
Learn from browsing in Classic and Smart Windows.

There is one extra wrinkle: disabling future learning does not automatically delete memories that already exist. Existing memories can still be used until they are removed from Smart Window's memory-management settings.

2. Web search

Smart Window can search the web when the assistant needs recent facts or information not available in the current page, tabs, history, or memories.

Mozilla announced an Exa partnership on August 18, 2026 for this search layer. Its support documentation says Smart Window can run a search through Exa, read relevant returned pages, and use those results to construct an answer.

That means a local LLM does not magically turn current web research into an offline operation. If a question needs the internet, the search step is still an online feature.

This is not necessarily a privacy failure; it is simply a different data path that should not be confused with the model endpoint.

3. Normal browser networking

Smart Window is still Firefox. Opening websites, loading scripts, images, fonts, analytics, APIs, videos, and third-party resources still requires the normal web connections those pages use.

Running the assistant model locally only affects the model call. It does not turn the browser itself into an offline application.

What stays local according to Mozilla?

Mozilla says several important pieces remain on the device.

Chats

Smart Window chat conversations are stored locally rather than synced as a cloud chat history.

Memories

The generated memory records are also stored locally. Mozilla says they are not currently synced across devices.

Private Window activity

Smart Window does not use activity or data from Private Windows to create memories.

Passwords, payments, unread email, and device files

Mozilla's current Smart Window documentation says the assistant cannot access:

passwords;
payment information;
unread emails;
files on the device.

That makes Smart Window different from a full desktop agent. It can reason about pages, referenced tabs, history, and memories, but Mozilla explicitly says it cannot independently click buttons, fill forms, complete purchases, book travel, change browser settings, or sign into sites on a user's behalf.

Built-in model vs remote custom model vs local model

There are now three very different ways to run the Smart Window assistant.

Setup	Main advantage	Main trade-off	Best fit
Built-in Smart Window model	Easiest configuration; Mozilla manages the provider path	Assistant requests leave the device through Mozilla's proxy	People who want the feature with minimal setup
Remote custom endpoint	Choice of provider/model and account	Data goes to the chosen third-party endpoint under that provider's terms	Users who already pay for or trust another model API
Local Ollama/Lemonade endpoint	Model inference can stay on the machine	Requires local model setup, RAM/CPU/GPU resources, and may behave differently from Mozilla's supported choices	Users prioritizing local inference and model control

Mozilla says its built-in AI requests go through a Mozilla proxy before reaching the AI service. That proxy is designed so the provider does not see the user's IP address or a unique Firefox/computer identifier. Mozilla also says built-in model providers do not store conversation data and do not use it to train their models.

A custom remote endpoint is a different trust decision. If a user points Smart Window at a third-party API, that provider's own privacy, retention, pricing, logging, and security policies become relevant.

A localhost endpoint removes that external model provider from the assistant-inference path—but only that path.

A privacy-first Smart Window configuration

For someone who wants to experiment while minimizing external AI processing, a more complete setup looks like this:

Step 1: use a local model endpoint

In Firefox:

Settings → AI Controls → Smart Window settings → Assistant model → Custom: Use your own LLM

Then configure a trusted local OpenAI-compatible endpoint such as Ollama or Lemonade Server.

Verify that the endpoint actually points to localhost or another intentionally controlled machine. A "custom" model can also be remote; custom does not mean local.

Step 2: turn off memory learning

Disable both memory-learning sources if browsing history and Smart Window chats should not be sent to Mozilla for memory generation.

That means turning off learning from:

browsing activity;
Smart Window chats.

Step 3: delete existing memories if personalization is no longer wanted

Stopping new memory creation does not remove existing memory records.

Open Manage memories and delete individual entries or all memories if the goal is to stop personalized responses entirely.

Step 4: use Private Windows for activity that should stay outside Smart Window memory

Mozilla explicitly says Private Window activity is excluded from Smart Window's memory system.

This is a clearer boundary than relying on a vague assumption that a particular page will not be remembered.

Step 5: treat web search as a separate online choice

If a request needs recent information, Smart Window may use Exa. A localhost LLM cannot answer live-web questions without some network source.

When the privacy requirement is strict, distinguish between:

summarize this page with my local model;
search the live web and tell me what changed today.

The second task inherently requires an external data source unless a separate local index has already captured the information.

The bigger lesson: “local AI” is not one switch

AI products increasingly combine several components:

browser context
      ↓
retrieval / web search
      ↓
memory generation
      ↓
LLM inference
      ↓
local history / storage

Moving the LLM box onto the user's machine improves one part of that pipeline.

It does not automatically move retrieval, sync, telemetry, memory processing, authentication, or ordinary network traffic onto the machine too.

This is why the phrase “runs locally” should always trigger a second question:

Which part runs locally?

For Firefox Smart Window, Mozilla now gives users unusually explicit controls over several of those layers. But the controls still have to be configured separately.

Is Smart Window a local alternative to AI browsers?

It can be, for a narrower kind of task.

Smart Window's assistant can:

summarize pages;
compare information across selected tabs;
answer questions about browsing history;
organize or close tabs through the browser's own supported controls;
use optional memories for personalization;
search the web when external information is needed.

But Mozilla says the assistant cannot act independently on websites. It cannot fill checkout forms, buy something, book a hotel, or sign into an account.

That makes the local-model option more interesting for browser-side reading, comparison, recall, and organization than for autonomous web automation.

If the desired job is "help me understand these six tabs without sending them to a general-purpose remote model," the local endpoint is a meaningful option.

If the desired job is "log into this site, click through five screens, and complete a transaction," Smart Window is currently the wrong category of tool.

What is confirmed—and what is not

Confirmed as of August 23, 2026

Smart Window is an early beta for Firefox desktop.
It can use Mozilla-provided model choices or a custom OpenAI-compatible endpoint.
Mozilla documents local setups using Ollama and Lemonade Server.
Smart Window chats and resulting memories are stored locally.
Memory creation can involve temporary processing on Mozilla servers.
Web search can use Exa.
Private Window activity is excluded from Smart Window memories.
Smart Window can be blocked through Firefox AI Controls.
Mozilla says custom models may not work exactly as expected.

Not established by the current documentation

Mozilla's public support pages do not justify the blanket claim that selecting a local model makes every Smart Window feature fully offline.

They also do not promise that every local model will produce the same quality, tool behavior, latency, or compatibility as Mozilla's built-in choices.

That uncertainty is why the safest description is local assistant inference with separately configurable online features, not "fully local Firefox AI."

A 60-second decision checklist

Before turning on Smart Window, answer these five questions:

Do I want model inference local? If yes, use a trusted localhost endpoint.
Do I want Firefox to create memories from browsing or chats? If no, disable both learning sources.
Do I need live web search? If yes, expect an online retrieval path.
Do I share this computer? If yes, remember that chats and memories are local but can still be visible to someone with access to the same profile/device.
Do I actually want an autonomous browser agent? If yes, Smart Window currently is not one.

That checklist is more useful than treating "AI on" and "AI off" as the only two privacy settings.

Conclusion

Firefox Smart Window's local-model support is real: a user can point the assistant at Ollama, Lemonade Server, or another OpenAI-compatible endpoint running locally.

The important caveat is architectural. Smart Window's model, web search, memory generation, and local storage are separate systems with separate data flows.

For the most local setup, use a localhost model, disable memory learning, remove old memories if personalization is not wanted, keep sensitive browsing in Private Windows, and remember that live web search is still an online operation.

The right question is no longer "Does Firefox AI run locally?" It is "Which Smart Window layer am I willing to let leave the device?"

Sources

Checked August 23, 2026:

Written and reviewed by /lico

Just writing down my thoughts, interests, and the things I learn along the way.