[avaer]

It works, I've shipped this as a "local inference"/poor person's ollama for low-end llm tasks like search. The main win is that it's free and privacy preserving, and (mostly) transparent to users in that they don't have to do anything, which is great for giving non-technical users local inference without making them do scary native things.

But keep in mind the actual experience for users is not great; the model download is orders of magnitude greater than downloading the browser itself, and something that needs to happen before you get your first token back. That's unfixable until operating systems start reliably shipping their own prebaked models that an API like this could plug into.

[Yokohiii]

> That's unfixable until operating systems start reliably shipping their own prebaked models that an API like this could plug into.

Maybe the next big thing will be some software subscription premium offers with a bunch of 5090s as an extra.

[Tepix]

It's a one-time download shared by all websites that use the Prompt API.

What's a bigger issue is that the models on most standard PCs are both tiny and slow. I was going to try using the Prompt API to change the text of (infocom) text adventures on the fly. But for many PCs, this will currently be too slow to be feasible.

[zozbot234]

> But keep in mind the actual experience for users is not great; the model download is orders of magnitude greater than downloading the browser itself, and something that needs to happen before you get your first token back.

With MoE models, you could fetch expert layers from the network on demand by issuing HTTP range queries for the corresponding offset, similar to how bittorrent downloads file chunks from multiple hosts. You'd still have to download shared layers, but time to first token would now be proportional to active-size rather than total-size. Of course this wouldn't be totally "offline" inference anymore, but for a web browser feature that's not a key consideration.

[NitpickLawyer]

> With MoE models, you could fetch expert layers from the network on demand

This is a common misconception, probably due to the unfortunate naming. Expert layers are not "expert" at any particular subject, and active-size only refers to the activated layers per token. You'd still need all (or most of all) the layers for any particular query, even if some layers have a very low chance of being activated.

All in all, you'd be better off with lazy loading the entire model, at least you'd know you have the capability to run inference from then on.

[zozbot234]

Ultimately it would amount to lazy-loading the model, but the parameters themselves would be fetched from the network as needed, which still decreases time-to-first-token. It's true that "expert" choices will span most of the model, regardless of any particular "subject" or "topic" choice, but if we simply care about time-to-first-token it's still a viable strategy.

[bavell]

Perhaps you could generate a few tokens before the entire model is downloaded, but since every token takes a potentially different "path" through an MoE model, you'd still need to wait for the entire download before getting deeper than a handful of tokens... which is not really a UX improvement imo.

[zozbot234]

Even at its worst, it's a minor UX improvement compared to having to download everything prior to getting to the first token. Ultimately we will complete the download, but we can still pick the best priority so that the first handful of tokens goes through.

[paganel]

> operating systems start reliably shipping their own prebaked models

Here's to hoping that that dystopia will never happen.

[aembleton]

Would it be less dystopian for Operating Systems to ship with their own browser that ships with their own models? Or do you find the current situation where Operating Systems ship with browsers dystopian?

[paganel]

I find the very idea of this AI thing making its way through like a virus onto our computer systems (either in centralised form, as it mostly happens now, or installed locally, like this article writes about) quite dystopian. On the other hand I do not find the idea of the Internet browser as dystopian (even though a data and corporate behemoth like Alphabet being the entity behind Chrome is indeed dystopian, I agree on that).

On a second thought you're on to something, maybe if we hadn't let the Internet browser take over our computer-lives as much as it did in the last 20 or so years then Chrome (under its current manifestation, that is) wouldn't have happened. At least we are now aware of the dangers awaiting in front of us when it comes to AI.

[halJordan]

Congratulations it already is here

[subhobroto]

> It works, I've shipped this as a "local inference"/poor person's ollama for low-end llm tasks like search

fantastic!

> the model download is orders of magnitude greater than downloading the browser itself, and something that needs to happen before you get your first token back

sure but does this mean the model is lazily downloaded? that is, if I used this and I am the first time the model was called, the user would be waiting until the model was downloaded at that point?

that sounds like a horrible user experience - maybe chrome reduces the confusion by showing a download dialog status or similar?

also, any idea what the on disk impact is?

[avaer]

The model download is lazy and cached, so it's a one-time cost presumably across all origins (I assume so since the alternative would be a trivial DoS waiting to happen).

So it's once per browser, not once per site.

You can track the download state yourself and pop whatever UI you want.

[tastroder]

chrome://on-device-internals reports "Model Name: v3Nano Version: 2025.06.30.1229 Folder size: 4,072.13 MiB" on a random Windows machine I just checked.

[subhobroto]

Thank You stranger! I would have assumed the size would vary based on whether your hardware supports the high-quality GPU backend (4 GB) or defaults to a smaller CPU-compatible version (3 GB) but the 22GB note on that page is really confusing. Even if it was including the model server where's the remaining 18GB going towards?

[danpalmer]

I'd imagine that the 22GB was decided through modelling various scenarios. For a start, it's not just a 4GB current model, it's 2x4GB to be able to update it without needing time when the computer is without a model, that's up to 8GB.

Then it's possible the model you get will scale with the CPU/GPU/RAM available, so if you have a 12GB GPU you probably get a better model, perhaps that's a 10-11GB model? At 2x that's 22GB.

Then consider that a machine is not static, GPUs/hardware come and go, VRAM allocation in integrated graphics changes, etc. You end up with just needing to pick a number and not confuse users.

[domenicd]

(Former Chrome built-in AI team member here.)

This is part of it, and also we just didn't want to use up the last of the user's disk space! It's disrespectful to use up 3 GB if the user only has 4 GB left; it's sketchy if the user only has 10 GB. At 22 GB, we felt there was more room to breathe.

One could argue that users should have more agency and transparency into these decisions, and for power users I agree... some kind of neato model management UI in chrome://settings would have been cool. But 99% of users would never see that, so I don't think it ever got built.

[why_is_it_good]

> Storage: At least 22 GB of free space on the volume that contains your Chrome profile.

[dotancohen]

Yes, but that is then followed by:

  > Built-in models should be significantly smaller. The exact size may vary slightly with updates.

[taejavu]

Lmao and here I am still staunchly treating Blazor’s 2MB runtime as a deal-breaker

[dotancohen]

Emacs had long ago exceeded eight megs!

[qingcharles]

If it doesn't fit on a floppy...!

[subhobroto]

> `> Storage: At least 22 GB of free space on the volume that contains your Chrome profile.`

Yes, I can read and comprehend English and you should assume I read the page. Because of the "At least" wording, I was curious what a person who has actually used the feature has noticed, aka, learning from people who have actually done it already.

[jfoster]

Doesn't sound great, but consider how much better this is than every webpage trying to load their own models.

If it turns out useful enough I'm sure browsers will just start including it as (perhaps optional?) part of installation.

[2ndorderthought]

Is it actually privacy preserving? Chrome mostly exists to extract all the information from a user it can without immediately getting a lawsuit of greater penalty than what is gained through ads, military contracts, etc. Android isn't too far off either. I would welcome any alternative to this. I can see applications for this being things like "while device is at rest and charging summarize all of the users recent text communications" or whatever else as a legal loop hole for wiretap laws

[gruez]

>I can see applications for this being things like "while device is at rest and charging summarize all of the users recent text communications" or whatever else as a legal loop hole for wiretap laws

This just exposes an API for sites to use. If they wanted to do the types of spying you're cynically suggesting, they could just add it without an API and you'd be none the wiser. Chrome contains closed source components so you wouldn't even know.

[masfuerte]

Do you think no-one would notice that the Chrome download was 20GB larger?

[gruez]

Who says they'll be using the 20GB model? You'd hardly need frontier level intelligence to detect CSAM or ad keywords. Moreover, it's downloaded on first use, not bundled with the browser, so you won't really notice unless you're checking the chrome user data directory, but that also contains caches and other site data you'd likely chalk it up to random sites.

[2ndorderthought]

It's a lot easier to hide the language they need in a EULA for a feature like this than it would be elsewhere.

I appreciate you feel this is a cynical take. But have you seen the class action lawsuits against Google over the last 5 years? They exceed a billion dollars as far as I can remember and they are for more blatant things than this.

[gruez]

>It's a lot easier to hide the language they need in a EULA for a feature like this than it would be elsewhere.

Why would adding a ML API or library require an EULA change?

[2ndorderthought]

So they don't get sued when the LLM inevitably instructs someone to drink bleach, hurt a stranger, delete the production db(I guess vibe coders say this is a good thing now though?), etc.

[kbx]

Hey, I'm the Chrome PM for the built-in AI APIs. I wanted to jump in on the privacy concern mentioned here.

It’s a totally valid question, and transparency is the only way this can work. On-device processing is an important core design goal of these APIs.

There are NO logs of the input / output interactions sent to any server, not even for training purposes. The only metrics we have are on performance, stability, and other generic API usage signals like any other APIs. These are all controlled by existing user preferences in Chrome.

[userbinator]

This comment sure didn't age well: https://news.ycombinator.com/item?id=48050964

[2ndorderthought]

Do you know if the android team plans to have access to this tool or if there is a plan for Android or chrome to integrate like this? Or maybe better, will other Google teams plan to use this tool to your knowledge to extract more user information? Surely this could change but it's always good to know.

Forgive my untrusting nature but I've been burned through several Google abuses of trust. I guess it doesn't matter much to me anymore I will never use chrome again, but maybe someone else is in the fence.