Breaking

Google Launches EmbeddingGemma 2 as a 740M-Parameter Open Multimodal Embedding Model

October 6, 2026 at 13:26 EDT

Foundation Models
Open Source
Software Dev & Coding

Google has released EmbeddingGemma 2, a 740-million-parameter open-weight embedding model designed to map text, code, images, video and audio into one shared vector space for local search and retrieval applications. The model is released under Apache 2.0 and is built for on-device use, with modular encoders that let developers load a smaller text-only configuration or the full multimodal stack.

Oct. 6, 2026 · Google

EmbeddingGemma 2 goes multimodal

A 740M-parameter open embedding model maps text, code, images, video and audio into one shared retrieval space for local apps.

740M

Full multimodal parameters

8,192

Token context window

768

Native vector dimensions

78.68

MTEB Code score

Storage compression becomes visual

768d full vector

128d truncation

Matryoshka Representation Learning lets developers shorten vectors from 768 dimensions to 512, 256 or 128 dimensions, reducing storage while accepting a quality tradeoff.

Load what is needed

270M text-only, 440M with vision, 570M with audio, or 740M full multimodal.

Local-first use cases

Media search, codebase retrieval, private RAG and zero-shot routing can run on device.

Configuration matters

128d vectors suit text best, dimensions must match, and float16 should be avoided.

Continue reading

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