Thanks Google, Embedding Gemma 2 is a big deal 🫶 Multimodal embeddings (text, images, video, audio) can run on-device in the browser. No server, no API key, ~20–70 ms per query on WebGPU. Everything stays on your machine (ofc). Now go build (new) things with it 🚀

Try the WebGPU demo here (credits to @xenovacom):

I did the experiment with Jev + memes templates. And it's hilarious haha

@nicdunz Does it work with legal adult content?

@grok is this a multi-vector or single-vector embedding?

We actually went deeper on this here:

on device means somebody finally stops paying per query

Embedding Gemma seems quite well timed for a project of mine. Also huge congratulations to the team!

embedding gemma 2 at 20-70ms in the browser with no key. cheap and small. our local litellm pool already takes that bill line to 0/month, and the thing that broke first when we moved models local was the context window, a 27b fits in 8gb at 4 bits just fine.

How big is the model download before the first query?

read this

Ahh ok

Gemini Embedding 2

I'd use this to search a folder of app screenshots. Does the demo keep the index between visits, or rebuild it each time?