Firefox Smart Window runs Mistral Small 4 — and you can swap it

Browser vendors have spent the last year welding one model into the chrome and calling it a feature. Mozilla went the other direction on September 16: Firefox Smart Window, its beta browsing assistant, now runs on Mistral Small 4 — and Mozilla says the model picker still works. That second half is the story. A default model is a business decision somebody made for you, and Mozilla is one of the few vendors that will tell you where the switch is.

What actually happened

Mozilla announced the partnership on September 16. Smart Window beta — the assistant that works across your open tabs, picks up a search trail you abandoned, and surfaces the source behind an answer — is now powered by Mistral Small 4 in the US, Canada and France. France is the first market outside North America, and it ships with French-language support rather than an English assistant with a translation layer bolted on. The UK and Germany are named for later this year.

Mozilla's own framing is blunt: across every market where Smart Window beta runs, Firefox users can still choose from other AI models. The stated goal is to keep people from "being chained to one company's self-serving pipeline."

Mistral's side of the announcement adds the part that matters contractually: conversations are not saved on Mozilla's servers by default, and partners including Mistral agree to zero data retention. Mistral lists four priorities for the deal — open-weight distribution, models tuned for regional languages and dialects, user control over privacy, and sovereign AI available to someone other than enterprise buyers.

Neither post publishes a parameter count or benchmark table for Small 4. Numbers floating around aggregator write-ups do not appear in either primary announcement, so treat them as unconfirmed.

Why Firefox Smart Window matters for your business

Two things here are worth stealing, and neither requires you to use Firefox.

The first is the model picker. Most AI features you buy this year will ship with a model already chosen. That choice sets your per-token cost, your latency, and your data path — and it can change under you in a version bump. When you build or buy, the question is not "which model" but "what happens when we want a different one." Build the swap in on day one or you will pay someone to retrofit it in eighteen months.

The second is zero data retention as a written term, not a marketing line. If you roll a browser assistant out to staff, every page it reads is your data leaving the building: customer records in a CRM tab, a pricing sheet, an invoice. "We don't train on it" is not the same promise as "we don't keep it." Ask for the retention clause in writing before the rollout, not after.

The regional-language claim is also testable, and you should test it. If you serve Spanish-speaking customers in Phoenix, a model advertised as tuned for regional dialects either handles the way your customers actually write or it does not. Twenty real messages will tell you in an afternoon.

The honest caveat: this is a beta in three markets, and a model picker only buys you portability if the thing you built around it is portable too.

Sources: Mozilla Blog, Mistral AI.