Google EmbeddingGemma 2
In brief
Google released EmbeddingGemma 2, a 740M-parameter open embedding model on the Gemma 4 architecture. It maps code, images, video, and audio into one unified space on edge devices for private, offline multimodal search.
How it works
Built on the Gemma 4 architecture, the 740-million-parameter model encodes text, code, images, video, and audio into a single unified embedding space. Its modular design reduces local vector storage requirements by up to 6x. Developers can run it directly on everyday devices to power offline search, such as finding a video clip from a voice memo or searching audio with text.
Developers building private, offline multimodal search applications for edge devices.
Availability
Sources
1 merged into this cardMore from Google

Gemini chat history and memory import
Google added tools to import chat history, context, and preferences from other AI assistants into Gemini. Users can bring over past conversations and tailored instructions without starting from scratch.

Tool Retrieval in the Gemini API
Gemini's Interactions API now supports on-demand loading of tool parameter schemas. Agents with large tool libraries can keep schemas out of context until needed, cutting input tokens, latency, and selection errors.
