Firefox Smart Window can send assistant requests to an Ollama model running on the same computer. The connection uses a model name and a local API address, so you do not need to buy an API key for the local model.
This guide takes you through checking access, preparing Ollama, connecting Firefox and checking a response against its source. It also explains a detail worth sorting out before you use personal tabs: local model inference does not make every Smart Window feature offline.
Check whether Smart Window is available first
As of 16 September 2026, Mozilla describes Smart Window as an early desktop beta rolling out gradually in the United States and Canada from Firefox 150, and France from Firefox 155. It requires a Mozilla account. Australia is not listed in the current rollout, and installing the latest Firefox does not guarantee access. Check the current availability information before downloading a model specifically for this feature.
Look for New Smart Window in Firefox’s menu. The separate AI chatbot sidebar is a different feature. If Smart Window is missing, stop here and check the rollout or waitlist; changing your Ollama settings will not make a missing Firefox feature appear.
Mozilla and Mistral’s 16 September announcement brings fresh attention to Smart Window. The custom-model connection described below already existed before that announcement. Selecting a hosted Mistral model and connecting a local Ollama model are different configurations.
What you need
- Firefox desktop with Smart Window enabled for your account and region.
- Ollama installed on the same computer.
- A downloaded local model that your computer can run.
- A short, public webpage for your first check.
Mozilla uses qwen3.5:4b in its Ollama example. The model listing shows an approximately 3.4 GB download. That is not a promise that 3.4 GB of free RAM or VRAM is enough: the runtime and conversation context need memory too. If choosing a model is the difficult part, our llmfit guide explains how to assess a local model against your hardware.
1. Download the model and check Ollama
Install Ollama from its official download page and leave the application or service running. Open PowerShell on Windows, or a terminal on macOS or Linux, and run:
ollama --version
ollama pull qwen3.5:4b
ollama listThe pull command downloads the model. Check that qwen3.5:4b appears in the list, then open a terminal conversation:
ollama run qwen3.5:4bTry a small request such as Explain what a browser tab is in one sentence. The purpose is to establish that the model can load and respond before Firefox is involved. Enter /bye when finished. Exiting this conversation does not mean you should quit the Ollama background service.
If you already use another local model, keep its exact installed tag for the Firefox configuration. A model answering a terminal prompt does not establish that it supports every browser action. Mozilla warns that custom models may not behave as expected.
2. Check the compatible API address
Firefox needs Ollama’s OpenAI-compatible base address. This is a local compatibility interface; it does not require an OpenAI account.
In Windows PowerShell, check the available models with:
Invoke-RestMethod -Uri 'http://localhost:11434/v1/models'On macOS or Linux:
curl http://localhost:11434/v1/modelsLook for the model you downloaded. A response establishes that the local server is reachable, not that Firefox is connected yet. If the request is refused, start Ollama and repeat this check before changing anything in the browser.
For this same-computer setup, keep Ollama’s default loopback binding. You do not need to expose port 11434 to the internet, add router forwarding or allow every web origin.
3. Connect Firefox Smart Window to Ollama
Follow Mozilla’s custom-model instructions:
- Open Firefox Settings.
- Choose AI Controls, then Smart Window Settings.
- Under Assistant model, select Custom: Use your own LLM.
- Enter the values below and select Save.
| Field | Value for this example |
|---|---|
| Model name | qwen3.5:4b |
| Model endpoint | http://localhost:11434/v1 |
| API key | Not required for local Ollama |
Use the base address exactly. Do not replace it with /api/chat or append /chat/completions in this settings field. If you chose another model, use the exact name shown by ollama list.
4. Verify the connection with a public page
Open a new Smart Window and ask a short question. Then open a public help page and request a three-point summary of that page. Compare each point with the original text. Avoid private email, work systems and account pages while you are establishing how the configuration behaves.
A useful check has three parts:
- Configuration: the assistant settings still show your custom model and localhost endpoint.
- Execution: Ollama shows activity while the request is being answered. Run
ollama psto inspect loaded models, or consult its server log. - Result: the answer describes the page you supplied, and you can verify the details yourself.
A reply that sounds plausible is not sufficient proof of the selected processing route. Keep the settings and server activity checks alongside the answer check. Start with one page before adding more tabs or asking the assistant to change anything.
For a separate local workflow with explicitly configured tools, see our llama.cpp MCP setup guide. That is a different integration, not a requirement for Smart Window.
What stays local, and what may still use the network?
A local model endpoint controls where the assistant’s model request goes. It does not switch off other browser services. Mozilla’s Smart Window privacy notice distinguishes custom-endpoint requests from its separate feature processing.
| Component | What to check |
|---|---|
| Assistant model | A downloaded model at a localhost Ollama endpoint runs on your computer. A remote custom endpoint still sends requests elsewhere. |
| Memories | Local storage of memories does not mean local creation. Review the learning controls separately. |
| Web search | Smart Window can use its web-search service. This is separate from local inference. |
| Browser services | Sign-in, website requests and operational data processing remain separate considerations. |
For a setup that does not learn new memories from your activity, open the Memories section of Smart Window settings and turn off both Learn from chats in Smart Window and Learn from browsing in Classic and Smart Windows. Review existing entries through Manage memories; disabling new learning does not delete them. The per-chat memory switch controls use in that chat, not all future memory creation. Mozilla explains the distinction in its Memories guide.
Ollama also has its own optional local-only setting, OLLAMA_NO_CLOUD=1. It takes effect after restarting Ollama and disables Ollama’s cloud features. It does not disable Firefox’s services. See the Ollama FAQ for platform-specific configuration.
Troubleshooting
There is no Smart Window option
Check Mozilla’s current rollout, your Firefox version and account onboarding. Availability is the first question, not the model download. The standard AI chatbot sidebar does not prove Smart Window is enabled.
The connection is refused
Repeat the /v1/models check outside Firefox. If that fails, confirm the Ollama application or service is running. On Linux, ollama serve starts a server when one is not already running. An address-in-use error can mean the normal service already owns the port; do not repeatedly start extra instances.
Firefox reports a missing model
Compare the settings field with ollama list, including the part after the colon. Pull and run that exact tag. A similarly named model on a download page is not necessarily installed on your computer.
The API responds, but Firefox still fails
Recheck the selected custom assistant and the /v1 base address. Inspect the server log for the request and error. On Windows, Ollama documents %LOCALAPPDATA%\Ollama\server.log; on macOS, ~/.ollama/logs/server.log. Linux systemd installations can use journalctl -u ollama. The official troubleshooting page covers these locations.
If the error mentions an origin or HTTP 403, investigate that specific failure. Do not start by setting wildcard origins, disabling browser protections or opening the server to your network. Mozilla’s documented same-computer recipe does not require those changes.
Responses are slow or page summaries are poor
Try one short page and inspect ollama ps. Its processor column helps distinguish CPU execution from GPU offloading. Reduce the task size before trying a larger model or context. Successful setup and useful answers are separate checks: reject summaries that introduce facts absent from the page.
Frequently asked questions
Can I use this in Australia?
Australia is not listed in Mozilla’s Smart Window rollout documentation checked on 16 September 2026. Check the current availability page before following the browser steps. Ollama itself can still be used independently.
Do I need an API subscription?
Not for a downloaded local Ollama model. A remote provider or cloud model may have its own account and charges. Keep the localhost endpoint and installed local tag distinct from those services.
Is Firefox Smart Window with Ollama completely offline?
No blanket offline claim follows from this configuration. The model can run locally while browser features still use the network. Review Memories, search and Mozilla’s privacy notice before supplying sensitive material.
Can I switch back?
Return to Assistant model settings and choose a provider offered in your Firefox installation. That changes the assistant’s processing destination back to that service. You can also return to a Classic Window without removing your local model.
Sources and update note
Configuration and availability checked against official Mozilla and Ollama sources on 16 September 2026. Smart Window is a beta feature, so check the linked help pages if the controls or supported regions change.