Breaking

Google launches EmbeddingGemma 2 as an open multimodal embedding model for on-device AI

October 6, 2026 at 12:19 EDT

Foundation Models
Open Source
Software Dev & Coding

Google DeepMind launched EmbeddingGemma 2 on Oct. 6, 2026, extending its on-device embedding family from text into an open, natively multimodal model for text, code, images, video and audio. The release follows more than 20 million downloads of the first EmbeddingGemma and gives developers a 740M-parameter option for private cross-modal search and retrieval on local hardware.

Oct. 6, 2026 | Google DeepMind

EmbeddingGemma 2 moves multimodal search on device

An open 740M-parameter embedding model maps text, code, images, video and audio into one 768-dimensional space for local search, retrieval and classification.

20M+

EmbeddingGemma downloads before the launch

740M

Full multimodal parameter count

8,192

Token context window

Maximum vector storage reduction

Modular footprint

Each vertical stack shows parameter load for the active encoders.

270M

Text only

440M

Text + image

570M

Text + audio

740M

Full multimodal

A 4x longer context window

The shift from 2K to 8K tokens gives the model room for longer local documents, audio and video inputs.

EmbeddingGemma 1

EmbeddingGemma 2

What developers can use

Private media and document search
Codebase retrieval for local tools
Voice or image queries over stored files

What needs validation

Memory on smaller machines
Quality after heavy truncation
Prefix, dimension and precision settings

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.