EmbeddingGemma 2 maps text, images, video, audio, and code, running 100% locally on your device. At just 740M parameters, you can build powerful multimodal apps running entirely in your browser via WebGPU & Transformers.js. From multimodal search to cross-modal classification and clustering, local embeddings unlock endless possibilities.

Credit to @xenovacom for building this demo

Embedding text, image, audio, and code into a unified vector space at just 740M on-device with WebGPU is massive for privacy-first apps. The latency and offline ergonomics will unlock so many browser-native workflows.

Five modalities in one space, on-device, matters most where data can't leave the device. Curious about sequences: for timestamped events (a card's payments, app logs), does it embed each item, or can it pool a whole sequence into one vector?

Entirely in the browser is the part that gets me

Now we’re cooking 🧑‍🍳

multimodal embeddings running fully in the browser at 740m parameters? this could make search and classification apps much easier to build without sending user data to a server. love the local-first approach.

i also made something out of this