Key points
- Browser AI runs right where a web page is open, with nothing installed.
- Since the computation finishes inside your own device, the photo or text you put in doesn't need to be sent to a server. That holds unless the page was built to send it anyway.
- In exchange, the first visit spends time downloading material. Once it's downloaded, later visits are fast.
- What it can handle is capped by what your own device can carry. A large model just doesn't fit.
- Hand someone a single address and that's all it takes, which fits a classroom or a demo especially well (as long as the devices can handle it).
The analogy
You've probably ducked into one of those unmanned photo booths on the street. No appointment, no equipment needed. Pull the curtain, press the button, and in that one small space the shot gets taken and printed, and a few minutes later you walk out with a photo. Nothing gets sent off anywhere, and your face never leaves that booth.
In exchange, it can only do what the booth is built for. The backgrounds are a handful of fixed choices, and a large group photo won't fit in the frame to begin with. Sometimes there's a short wait for the screen to power up and get ready, too.
Opening a web page and having AI run right there inside it works the same way as that booth. This is browser AI.
In detail
What "nothing installed" means
Open the address and that's it. No program to download, no admin rights, no account needed. That means it works even on a computer where installing software is locked down — a school or company machine — which makes it especially convenient for a class or an in-house demo.
Handing over one address is enough for someone else to open the same page. When the people who built it update the page, the next person to open it automatically gets the new version. The hassle of keeping versions in sync disappears.
Windows, Mac, tablet — there's no need to build a separate version for each device. That said, unless the model is a lightweight one, it needs a recent browser and plenty of memory, so depending on the device it can run slowly or not open at all. Not having to build separate versions is a big win for whoever's building it.
What you put in doesn't leave
In browser AI, the computation happens inside your own device. A face caught on a webcam, a document you dragged in, a sentence you typed — none of it has to be sent anywhere for the computation to work.
That's a relief when you're handling material you'd rather not send out — blurring a face in a photo, sorting through a document that has personal information in it, the kind of task that turns risky the moment it goes to a server.
Once the material has finished downloading, it can keep working even with the internet cut off. Watching the conversation on screen keep going after the subway loses signal makes the character of this approach obvious at a glance.
The first wait, and what your device has to carry
It isn't free. The first time you open the page, it has to download the model's material, and a small model, like one that finds faces in a photo, needs only a few megabytes, but a conversational language model runs from a few hundred megabytes to a few gigabytes. That stretch shapes the first impression.
Once downloaded, the browser keeps that material around and reuses it on the next visit, so the wait drops sharply after the first time. Clear your storage or open it in a different browser, though, and you're back to downloading from scratch.
How fast it runs comes down to your device. Modern browsers can borrow the power of graphics hardware for the calculation, which is much faster than it used to be, but a large model still just doesn't fit in memory. Run it for a long stretch and the device heats up, with the battery draining noticeably.
Sorting out similar-sounding terms
Look only at "runs on the device" and this sounds the same as on-device AI. The line that actually separates them is the fence. Browser AI stays inside the fence of a web page, borrowing only as much of the device's power as the browser allows. On-device AI reaches the chip built into the device directly and is attached to something like a camera app or the keyboard.
There's also the separate approach of installing a program on your own computer to run a large model. That one needs installing and setting up, but in exchange it can handle something much bigger.
Choosing between the three is simple. For letting anyone open it through a link, pick browser AI; for something attached to a device feature and always on, on-device; for running something large properly, install it.
More precisely
Heavy computation became possible inside a browser because of two things: a channel that lets a web page reach the graphics hardware, and a runtime that executes fast inside the browser. Model material gets stored in the browser's own storage and reused on the next visit. Browsers are also starting to carry a small model of their own that a page can borrow.
The analogy breaks down in one place. A photo booth's equipment sits there permanently, but browser AI's material is only temporarily perched on your device — clear your storage and it's gone. And when a booth's curtain is closed, nobody knows what happens inside, but the computation on a page can be built, if whoever made it chooses to, to also send the result to a server. The claim that nothing leaves your device only holds for a page built to send nothing to a server.
Try it yourself
- WebLLM Chat (a chat AI running in your browser) ailearn.space After the first visit finishes downloading its material, check whether the conversation still works with the internet disconnected
- TensorFlow Playground ailearn.space The whole training process happens right inside your browser — there's nothing to install
Common misconceptions
It's easy to think every AI used in a browser works the same way, but actually most web services send your input to a server to compute it, and only browser AI downloads the model and computes inside your own device. Which one you're getting differs from service to service.
It's easy to think anything running in a browser is automatically slow, but actually a well-trimmed small model can respond faster than a round trip to a server.
It's easy to think it works the same on every device, but actually it can fail to load at all depending on the device's memory and the browser.
One-line summary
Browser AI runs like an unmanned photo booth, inside a page with nothing installed, and the price of not having to send your material out is a first wait and a hard cap set by your device's power.
Spotted an error or have a better analogy? Suggest an edit · Last updated 2026-10-18