EmbeddingGemma 2

Google's small open model that makes text, images, audio, and video searchable on a phone or laptop, for developers building on-device search.

Maps text, code, images, audio, and video into one shared space, so a single query can match across formats. The first version handled text only.
740M parameters on the Gemma 4 architecture. Text-only use needs as little as 270M parameters, with the vision (170M) and audio (300M) parts optional, about 191MB of RAM on a Pixel 11 Pro.
Handles 8K tokens per input, four times the first version, which is enough for about 5.5 minutes of audio, 29 images, or 58 video frames.
Flexible embedding sizes cut storage in a local database by up to 6x. Google also calls out a large jump in code search quality over the first version, aimed at indexing local codebases.
Open weights under the Apache 2.0 license on Hugging Face and Kaggle, with built-in support for LiteRT, MediaPipe, LangChain, and LlamaIndex.