Why doesn't HuggingFace create a browser extension for transformers js?

[AgustinRamiroDiaz]

Question

Given that models cannot be shared cross site (due to web browser security), using different apps with transformers.js implies downloading a model many times.

Why isn't there a browser extension like https://github.com/alexanderatallah/window.ai from HuggingFace in order to share models + also let the users pick the models they prefer?

I envision a web browser extension where users can manage the models they have, and then each web app built with transformers.js could connect to this extension as a provider

---

Google is doing something similar to this: they are testing out shipping Gemini Nano in Chrome, so then they have different AI APIs that use the model https://developer.chrome.com/docs/ai/built-in-apis . I know it's a different situation than the one from Hugging Face, since Google has more weight on browser APIs

[nico-martin]

Hi @AgustinRamiroDiaz, thanks for reaching out!

Just to set some context first: Transformers.js is primarily a building block, a core library for running ML models in the browser, rather than a product built on top of it. Building things like a model manager extension is very much in the "what others can build with this" space, and we intentionally keep our focus there.

Regarding Chrome's built-in AI APIs (Prompt API, etc.) and Gemini Nano: we actually love the direction Google is taking here. Having a model maintained by the browser and accessible through a unified, standardized API is a great idea, and we see it as complementary to what Transformers.js offers rather than competing with it. The key difference is that with Transformers.js you retain full control over which exact model you use. You're not limited to whatever the browser vendor ships.

On the browser extension idea: the main benefit I see would be acting as a shared cache layer. So instead of each origin downloading its own copy of a model, a single extension manages the downloads and serves them across sites. The good news is that anyone could build exactly this on top of Transformers.js, since you can provide a custom cache implementation.

And actually, this is something we're already actively exploring. There's a proposal for a Cross-Origin Storage API that would allow access to a global storage not bound to a single origin, which is essentially the browser-native solution to this problem. We've been working closely with the Chrome team on this and have already implemented an experimental caching strategy for it. You can try it out today with the Cross-Origin Storage Chrome extension and the following flag in Transformers.js:

env.experimental_useCrossOriginStorage = true;

This will attempt to use cross-origin storage if the extension is installed, and gracefully fall back to the standard Cache API if not.

⚠️ This is highly experimental. Expect things to change.

As a concrete example of all this coming together: there's a Prompt API polyfill that uses Transformers.js under the hood as the default backend and already has cross-origin storage enabled by default. It's a good reference for how this kind of shared-model architecture can work in practice today.

Hope that helps clarify where things stand.. and where they're heading!