Breaking

Google Releases EmbeddingGemma 2, an Open 740M-Parameter Multimodal Embedding Model for On-Device Search

October 12, 2026 at 11:57 EDT

Foundation Models
Open Source
Infra & Chips

Google DeepMind has released EmbeddingGemma 2, an open-source embedding model that unifies text, code, images, video, and audio into a single 768-dimensional vector space, enabling on-device semantic search without cloud connectivity. The 740-million-parameter model, released under the Apache 2.0 license on October 6, 2026, represents a significant step in making multimodal AI deployable on smartphones and edge devices.

OCTOBER 6, 2026 · GOOGLE DEEPMIND

EmbeddingGemma 2: Multimodal embeddings on every device

Google releases a 740M open-source embedding model that unifies text, code, images, video, and audio—all running on-device without cloud connectivity.

Key specifications

Total parameters

740M

Modular: 270M text/code only

Embedding space

768D

Compressible to 128D

Context window

8K tokens

4× previous version

Language coverage

100+

Via unified encoder

Performance comparison

68.46

English text

v2 vs v1 MTEB

+10.0%

Code search

Improvement

Multimodal innovation

Unified vector space enables direct comparison of text queries with images, video frames, and audio—without separate models for each modality.

On-device deployment

191 MB memory (text only) or 567 MB (full multimodal) enables production deployment on Pixel 11 Pro and other Android devices without cloud connectivity.

Continue reading

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