- 71comments
- 571comments
- 60comments
- 434comments
- 53comments
- 214comments
- 230comments
- 2comments
- —discuss
- 13comments
- 11comments
- 319comments
- 72comments
- 57comments
- 142comments
- 299comments
- 216comments
- 2comments
- 14comments
- 92comments
- 377comments
- 16comments
- 41comments
- 1comments
- 25comments
- 5comments
- 171comments
- 440comments
- 146comments
- 59comments
This sounds great, will have to try it out. Interested in the business value they gain from this, as this appears to be free.
I guess for mistral it's great marketing material.
Edit: Quite interested in the differences of local optimizations... that does allow for specific sensorship, but alas that's already possible in a more direct, reliable way.
I think that beyond marketing this is also a way to collect "AI browsing" trajectories. Not an expert but I think this is the kind of great quality data that can be useful when training LLMs.
What is "AI browsing"?
AI are excellent at ingesting tons of data and getting something relevant from what you asked. I ask it about linux behavior and it will dig through hundreds of patches & mails and will return me informations sourced with links to the patches, mails containing it's affirmations.
Seems like what Chrome has with the default built-in Gemini Nano model.
Hey at least "some" news/things from Mistral. Seems like ages ago when they launched vibe-code.
I don't think it is comparable. Nowhere in the news they say it is offline but IMO Gemini Nano is offline (?)
I'm not a Google fan but I never understood the outrage as it is a local model ... sure, you have to trust them, but if you use their browser you already do.
The outrage about Gemini Nano in Chrome was the surprise extra 4GB download plus their attempt to expose it as a non-standard browser API, IIRC.
Reading HN, I 100% understood “private” to be marketing copy for “local inference”.
How else could the information be kept private?
I begrudgingly suppose that there is precedent for the usage, like Apple’s Private Relay, but I really don’t think they are comparable.
I don't read this as offline ...
Pretty ironic then that here, Mozilla is specifically labeling their "we send it all to the cloud" service as "private."
This is an excellent use case for completely local, small model inference, yet for inexplicable reasons Mozilla wants to normalize uploading your entire private browsing history to a cloud.
These (Mistral's and Mozilla's) marketing pages aren't candid enough to clearly explain the difference between local and cloud inference, and that they're asking you to consent to enabling the latter. I'd call that the bare minimum of ethics. If you're in a position of authority over your less technical users, you are ethically obligated to give them the full picture of what are you doing and why they should consent.
(Aside to any Mozilla people who might be reading HN, your page here[0] has an oversight, it advertises this model as "Mistral Small 4" but the hyperlink is to OpenAI's model card for gpt-oss-120b).
[0] https://support.mozilla.org/en-US/kb/smart-window-models
I can't fathom how they'd consider "Private" something that uploads all the sensitive information in readable form to a third party service.
In fact, If I were asked to explain the opposite of private, I might give this kind of design as an example.
It's more un-fathomable if you recall Mozilla were the people who *pioneered* local browser ML as a privacy feature. Back in 2022—when Chrome was doing language translation by uploading to Google servers, and Mozilla introduced the (actually innovative!) idea of baking small ML models into the browser,
https://news.ycombinator.com/item?id=33792447 ("Firefox Translations: Translate websites in your browser without using the cloud (addons.mozilla.org)")
And look: they were 100% cognizant, then, of how vital local processing is to privacy:
https://blog.mozilla.org/en/mozilla/local-translation-add-on... ("Mozilla releases local machine translation tools as part of Project Bergamot")
Thanks for mentioning the mistake on the link, it's been reported internally.
The "Gemini 3.1 Flash Lite" also points incorrectly (?) to https://storage.googleapis.com/deepmind-media/Model-Cards/Ge...
The privacy policy for Smart Window[1] indicates that prompts are not only sent to an LLM provider (Mistral in this case), but also to Mozilla servers first.
Where they do use an on-device model is for the initial query classification:
[1] https://www.mozilla.org/en-US/privacy/smart-window/
Some years ago Firefox was the go to browser if you wanted to have "some" privacy in your browsing (together with uBlock origin). With news like this I m really wondering if my views are outdated and I need to to some good researching on maintaining some privacy in what I m browsing.
Mozilla has been in a bit of a recursive feedback loop death spiral for years now of: Browser loses market share -> try some weird thing -> very few people like or use it -> browser loses market share -> management says "oh shit we're losing market share we better try some weird thing"
I say this as a person that uses firefox with ublock origin 99.5% of the time. It's better in my opinion than Chrome. I can at least easily turn off the enabled by default crap features. Like the advertising and sponsored news links on the default new tab page.
This sentence is eye-opening:
Privacy-minded users who want to try this feature would need to create a totally separate browser profile to avoid data leaks.
Except ... this is not local. It's a remote service. Here is something that is:
Configure them. Open-webui needs a model (a local qwen3.8 does amazing), and you need to link them up. Playwright will just do stuff.
https://support.mozilla.org/en-US/kb/smart-window-byom
Also supports Lemonade and Ollama, which are.
Oohohohohoohohohoho Merry Christmas!!!
its up to us to make extensions that are so famous that they can be integrated into the browser functionality i guess
How small is this 'small model' inference?
Browsers especially are expected to operate on a very wide number of devices. It makes perfect sense to me to start off with a cloud driven approach and later allow people to configure it to run locally if they have the hardware to do so. Anyone with the hardware to do so is already signaling that they'd have the technical interest to make that change while having the default 'works for everyone' mode be one that doesn't require a powerful local device.
I say this as someone using a 6 year old computer and only upgraded to 8GB of ram when the latest WoW expansion was released.
I am not trying to defend Mozilla doing this, and I don't support sending data to cloud based services like this in a way that users won't understand.
But I also think that the state of the art in small LLM and user device capabilities aren't there yet to put a "good enough to be actually useful" local-only LLM as a prepackaged thing in a mass market distributed browser.
You don't want a browser that takes 10GB of extra RAM (on top of the memory hog that is having just 3 or 4 complex tabs open on its own already) and pegs your CPU at 99% usage for minutes at a time. And not in an era when mass market consumer laptops are still commonly 8GB or 16GB of total system RAM. Many of those with integrated-into-CPU onboard graphics (eg: not a gaming laptop with a discrete GPU on the PCI-E bus).
It'll be a catastrophe for laptop battery life, among other resource use problems. And an LLM that fits in under 8 to 10GB of RAM for CPU-only inference is not going to be nearly as capable as an off-device inference system.
I wish they had just done this with a very clear up front opt in (not enabled by default) thing that explains what Mistral is, that it's not some big American cloud company but a relatively small startup in France, and that your prompts/LLM interactions will go to their servers. And some documentation on how it will be handled/stored in a supposedly trustworthy manner.
Is it really ? LLM take lot of ram and drain battery. People run Firefox on low end computer.
Indeed. Realistically a 'capable' small local LLM, even one that's definitely not as good as externally hosted ones will require a single 16GB GPU and access to basically all of the RAM on the GPU. That's not something people running Firefox on a $500 laptop with 8 or 16GB of total system RAM and a CPU-integrated basic graphics system have to spare.
I certainly agree that this option would be nice to have, but running firefox without an ad blocker on the modern web will ALREADY spin up my fans just rendering all the garbage, and I have a decent laptop.
You can't expect the median Firefox-running device to be able to run a capable enough local model effectively.
Just gonna go ahead and adopt the phrase "spin up my fans" to describe pissing me off.
zdr is good, really confidential inference with attestation would be better. the same as a confidential vm TEE which confirms the integrity and privacy.
there are good providers available, i think it should become the standard.
two high quality companies: near.ai tinfoil.sh
Didn’t OpenAI tried to build even a new AI browser and failed? I honestly don’t need AI in my browser and would not let an agent to access my login credentials/cookies/saved payment methods for various sites.
It's pretty nice for tedious tasks. I tried to have it look through job listings based on my criteria. It burned through my 5h Claude quota, but what I got was pretty impressive.
I agree on the privacy concerns but this could be useful with a local model, I think.
For me it’s not only privacy, I just don’t trust an agent to use my credentials without doing any destructive task because unlike an MCP, I cannot enforce any permission model on the browser
this is chatgpt, web gui from mistral built-in Firefox, not for tech savvy users. I think anyone that will complain about this knows very well that you can just give your harness control over firefox through puppeteer, which is much more powerful. could firefox itself become a harness and speak openai api responses? maybe yeah, but why?
Related, but hardly activity:
"Firefox Smart Window" 21.apr.2026 https://news.ycombinator.com/item?id=47852016
"Firefox Smart Window" 18.aug.2026 https://news.ycombinator.com/item?id=49345179
"Firefox and Exa Partnership" 18.aug.2026 https://news.ycombinator.com/item?id=49346066
"Private"
Guess we just using buzzwords for marketing and have no morality or principles anymore. Bad timeline and people you should just not trust.
All previous similar issues got hand waved with "Mozilla has to make money somehow". I am afraid one day remaining users will decide that Chromium based browsers provide a better experience in most use cases, and now have same privacy guarantees as Firefox. Looking at market share maybe that had already happened and pragmatism won over ideology, and we will all suffer the consequences of such monoculture.
Glad to use ~~Unmozilled Firefox~~ Librewolf
Waterfox also works well!
I remember using Waterfox in the past, I think because they were keeping an old UI. Is that the draw over Librewolf?
For me the draw is that Waterfox has partnerships in order to provide funding so if/when Mozilla truly loses the thread (which some may argue they have) WF can keep going on its own, whereas the Librewolf devs have stated that if push comes to shove they can't maintain an entire browser themselves.
Mozilla lost the tech race and with it a lot of users. Firing a lot of their talented developers didn't help. Now Mozilla is giving up the only reason why they kept a core audience of privacy focused users by sending our browsing data to third party servers for inference just at the time when small local models are getting good enough for many/most use cases like summarization or translation. Chrome provides an API for local LLMs. Apple is trying to run as much interference locally as possible. Mozilla is worse than that. Claiming private browsing while sending data to third parties.
I don't care what their data retention policies are, I don't want my data to be sent to others full stop.
Private means it's mine, it's under my control. Handing it to third parties is not private.
Please don't.
Would be very cool if you could type a long query and the model would just build an advanced google search query using what you typed. You could even ship a tiny model in the browser itself that does that.
Something like: "blog posts which talk about using sending emails to yourself as a notes taking mechanism".
The output could look like: "note taking" OR "notes" ("emailing yourself" OR "email to myself" OR "email yourself") (inurl:blog OR inurl:posts OR site:medium.com OR site:substack.com)
It wouldn't be cool because it would still hand your data to Google.
I assume Google was just a placeholder for a search engine in the example. And you'll surely need some service to search the web.
It could anonymize the data, at least.
I don't know about that. When I hear people reference "Google" it seems to me they do mean Google. Similar with "Claude".
People still use “Google” as a synonym for “online search.”
We’ve got the whole post to parse, though.
The idea is obviously applicable to other search engines.
That’s a great idea, and would be a much better feature than the unimaginative things Mozilla has been implementing. It’s so disappointing.
I think that's roughly what duck.ai is doing? If I type that in there I see it generating some search queries and then generating a list of links based on the different result pages.
I’m fine with cloud features, but I’d like to know what stays on my device and what gets sent out. “Private” is a bit vague.
I cannot in a million years imagine using these features. Just be more organised.
I understand why they built this. They probably should and this is probably the right way to do it.
And also yes, I’ve tried Dia and OpenAI’s thing and I can’t for the life of me figure out what to do with the feature.
And then it dawns on me what I really want is a modernized bookmarks system…
yeah ?
Not sure if it's a locally run model or in the cloud. If it's the former they don't explain the size on disk and more importantly on memory. I know that google is using local models on the pixel phones, so it's doable. If it's the later, the "Mozilla don't keep the data" is hypocritical and I'm perplexed by the cost associated.
In any case as a french, I prefer my personnal data be exploited by a french company :D
Edit: after reading the comments, it is not a local AI so "Mozilla don't keeep the data and don't train on them" is really hypocritical if they cannot confirm that it's the same on mistral side. And there is still the question of the cost.
It's interesting to see Firefox attempt to create slightly more privacy-focused cloud inference infra (assuming you can trust that they adhere to their own policies and don't have bugs, and that their partners adhere to their contractual obligations and don't have bugs). Unfortunately that's still a lot of trust that's pretty much impossible for the end-user to verify, but at least it's something? Arguably it's better than the alternative of directly trusting AI companies that have shown absolutely no regard for other laws like copyright, but the marketing of this as ultra-private still leaves a bad taste. I wish this announcement clarified the cloud aspects directly instead of needing to browse through two layers of links to find how it actually works on a privacy notice.
Fuck off with this AI bullshit.