Breaking

Google DeepMind Introduces 740M EmbeddingGemma 2 for On-Device Multimodal Search

October 6, 2026 at 16:18 EDT

Foundation Models
Open Source
Software Dev & Coding

Google DeepMind introduced EmbeddingGemma 2 on Oct. 6, 2026, a 740M-parameter open multimodal embedding model that maps text, code, images, video and audio into one 768-dimensional vector space for local search and retrieval. The release extends the text-only EmbeddingGemma line into on-device media search, RAG, classification and clustering, with the model published under Apache 2.0. EmbeddingGemma 2 model card

Oct. 6, 2026 · Google DeepMind

EmbeddingGemma 2 moves multimodal retrieval onto devices

The 740M-parameter open model maps text, code, images, video and audio into one shared vector space for local search, RAG and routing.

740M

Full multimodal parameters

Input modalities in one space

8,192

Shared context tokens

37.3 ms

Reported visual embedding latency

Load only the encoders you need

Each column uses stacked blocks to show effective parameter size on the same scale.

270M

Text

440M

Text + vision

570M

Text + audio

740M

Full

Why it matters

One local model can replace separate captioning, speech transcription and text embedding steps for many retrieval workflows.

Deployment checks

Use task prefixes, keep vector dimensions matched, avoid float16 inference and validate 128d truncation for multimodal search.

The test shifts from benchmarks to local apps

Media searchPrivate RAGCode indexingZero-shot routing

Continue reading

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