EmbeddingGemma 2

Open-weight 740M-parameter multimodal embedding model built on Gemma 4 that encodes text, images, audio, and video for on-device search.

Google · Data, training, and fine-tuning platforms (datasets, labeling, RAG/vector DBs, fine-tuning-as-a-service) · model

Published in Feed

Overview

EmbeddingGemma 2, released by Google on October 6, 2026, extends the original EmbeddingGemma beyond text into images, audio, and video, mapping all four modalities into a shared 768-dimensional embedding space. Built on the Gemma 4 architecture and released under Apache 2.0, it is designed for on-device, offline semantic search and retrieval so that local file and media search no longer depends on a cloud round trip.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

  • 740M total parameters: 270M text-only core plus optional 170M vision and 300M audio encoders
  • Unified 768-dimensional embedding space across text, image, audio, and video
  • Apache 2.0 open-weight license
  • ~191MB active RAM for text-only inference on a Google Pixel 11 Pro; ~567MB for the full multimodal suite
  • 9.92-point improvement on the MTEB Code benchmark over prior Gemma embedding models

Use cases

  • On-device semantic search across photos, audio, and video
  • Local codebase indexing and developer agent retrieval
  • Offline RAG pipelines without cloud dependency

Pricing

Free, open-weight model under Apache 2.0.

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released
6 Oct 2026
Record updated
13 Oct 2026
Record last checked
13 Oct 2026

Explore the Feed

Other entries filed under data, training, and fine-tuning platforms (datasets, labeling, rag/vector dbs, fine-tuning-as-a-service); inclusion does not imply equivalent features.

  • DigitalOcean Agent Data Services

    DigitalOcean's managed data layer for agents: ingestion flows, a catalog with vectors, and SQL or vector query, in private preview.

  • pplx-embed-v2-late

    Perplexity's late-interaction multimodal embedding models (0.6B and 9B) with 128-dim per-token vectors, open on Hugging Face.

  • AutoSynthData

    ServiceNow CoreAI's framework that turns an enterprise agent's capability gaps into targeted synthetic training data.

Browse the Feed