Urgent.News

What's breaking now, across thousands of outlets.

AI

EmbeddingGemma 2: an open, lightweight multimodal embedding model

EmbeddingGemma 2 marks a significant advancement in multimodal embedding models, expanding its capabilities beyond text to include code, images, video, and audio. This open-source model, released under a commercially permissive Apache 2.0 license, boasts 740 million parameters, making it well-suited for on-device inference. Built on the robust Gemma 4 architecture, EmbeddingGemma 2 offers enhanced performance in code processing, achieving a notable 9.92-point improvement over its predecessor, EmbeddingGemma.

This improvement, quantified in MTEB Code, is essential for local codebase indexing, semantic code search, and coding agent retrieval. The model's performance extends across various data types, setting new benchmarks in quality-per-parameter for sub-1B models and even surpassing some larger specialist models. EmbeddingGemma 2's introduction brings powerful capabilities directly to edge hardware, enhancing data privacy by enabling local processing.

This reduction in pipeline latency empowers developers to create cross-modal search and retrieval systems that operate entirely offline. When integrated with generative models such as Gemma 4, it facilitates on-device RAG pipelines capable of understanding complex multimodal data. The model's compatibility with Gemma 4's text tokenizer and audio encoder facilitates a unified pipeline with a reduced combined memory footprint.

Google AI Edge blog post provides detailed guidance on leveraging EmbeddingGemma 2 for on-device search and RAG systems, highlighting its seamless integration with LiteRT. The developers' close collaboration with industry partners ensures immediate applicability of EmbeddingGemma 2 across various development platforms, as detailed in the comprehensive developer guide, documentation, and inference/fine-tuning resources.

Written by urgent.news from Google DeepMind's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

This story

This is one outlet's version. Read the fullest account.

Every outlet that reported this, and when →

Read the original at deepmind.google →

More in AI

The next hurdle for AI agents: getting websites to let them in

Personal AI agents promise to shop, book flights, and make reservations for you. But deliberate blocks and anti-bot defenses are getting in the way, leaving consumers caught in the middle.

  • AI agents like Meta's Muse perform tasks like booking flights and ordering groceries.
  • Websites employ anti-bot measures that prevent AI agents from completing tasks.
  • Industry partners are developing an open standard for AI agents' communication with businesses.

More from Tuesday 6 October →