Post

Anecdata, but I am strongly strongly strongly positive about this API. The interoperability and the ToS issues do not detract from the idea here at all. Those are browser implementation specifics/details that in no way detract from the massive potential this has for the web.

I would like to see Google have model selection in Chrome. Unbundling the Gemini model from the web could alleviate both the inter-op & ToS concerns & would be good for users. But I'm not prepared to obstruct web progress basically upon the fact that Chrome is the biggest browser by far! Boo.

idk if “solving interoperability & ToS issues” or “building vendor consensus and community buy in around a feature before it ships” are exactly obstructing web progress … the absence of this api prevents nobody from using an LLM in a web app; they can just use a library for that right now!

Something like an LLM should not be built in to the web for the simple reason that it’s unpredictable, not exactly controllable, and will randomly misbehave sometimes, among other reasons. It barely deserves to be called computation- But wait- … … Can’t the same things be said about Javascript?

you are right about Javascript but unfortunately we are not ready for that conversation as a society

no, JavaScript has a standards process and a spec that defines how it works.

The ancient tomes of mystery! *1950's radio voice activates* "Undecipherable by all but the most fastidious adventurers, these scrolls impart forbidden knowledge that many never attain, like why 1 + '1' = '11' but '1'*1 =1"

Does the prompt API require user permission?

no

This is good and all, but do y'all sell that shirt? The kit is so cute!

Not this exact one (I got it custom printed) but there are some on mozilla-europe.myspreadshop.co.uk

@jakearchibald.com - "In Chrome when you create a LanguageModel it will either use the Google Cloud API or download a 4GB model to run locally". We don't use the Google Cloud API. If the model isn't available we will kick off a download, but the API just doesn't work without the modal locally.

Interesting! When if I uninstall the model, and chrome://on-device-internals/ confirms the folder size has gone from 4gb to 0MiB, the prompt API still works. Including in Incognito. What's going on here?

Even after full browser restart? Otherwise the model could still be loaded in memory?

is there a permission check on the users machine before downloading? is it possible that end users get llms downloaded to their machines without their knowledge?

WASM based LLM inference in the browser is possible for a while. Those often download models in that size. Any website can download up to ~10GB of data in localstorage IIRC, if enough space on disk is available. Or other blobs, like e.g. astronomical images than can easily be several GB in size.

I'm very grateful for people pushing back. I used the API, have some strong ideas how to make great things with it, but the legal aspects of it are broken as-is

Genuinely interested in your use-cases for it, and how you'd tackle the model being different from machine to machine & over time.

I'm interested in this for generating alt text on edge devices. And model being different device to device and over time is a nightmare and a half. I am p annoyed that they released this with gemini nano 3 (instead of waiting for the gemma4-based gemini nano 4), it's a waste of space rn.

I gave a talk on this with inspirational but overall useless demos. The idea was that the model would solve problems by generating code and the prompt was not tailored much beyond telling it that. The code would then run sandboxes in the page and be endowed with references necessary to do things.

you don't need to license Google's model, it's free and it's going to change - every model out there has a very limited life span; any LLM based app needs to be ready for different models the API itself may be half baked and definitely shouldn't be rushed through without consensus either way!

for that matter, their model is already outdated - other vendors already ship (free, open) models that are smaller, faster, and more intelligent the API may be rushed, but the system prompt and model lock-in concerns aren't real - or they are, but you're just describing the reality of using LLMs

Ever wondered why Google is shipping its own 4gb model rather than using the one already on the OS? For… consistency! Which is exactly the problem here.

Instance number 46285 of the Chromium team rushing a half-baked API out the door, explicitly avoiding standardization processes because they know their crap isn't compatible with the idea of a free and open web.

I find the hate mongering both inaccurate, distasteful, but most of all, quite low class. That just doesn't seem true to me, period, the team is doing enormous work to make the web much much better, and we need more people making good bets knt he web & trying. They don't deserve this rage.

That's a fair position. I should be clear that I don't mean to say that the Chromium contributors are a monolithic evil entity that only do bad things. They absolutely do good in much of their work, and we do greatly benefit from a lot of what they do.

Our full standards position: github.com/mozilla/stan...

seems like the system prompt is light-weight/basic and chrome exposes developer-level “system” prompts directly? i understand your concerns, but i am absolutely going to try this out.

i wonder if people are going to get as upset about google shoving this through the standards process as they got/still are about web components supposedly also shoved through by the googlers. guess we will have to wait 10 solid years to find out. as the vice president of web components, not a fan.

Wait did Jake jump to Mozilla?

I left Google in 2023 and joined Mozilla in 2025.