EmbeddingGemma 2 extends Google's lightweight embedding model beyond text, placing code, images, video, and audio in one shared embedding space. The original EmbeddingGemma has passed 20 million downloads, and the new version is built on the Gemma 4 architecture with 740 million parameters. For developers building on-device search and privacy-first retrieval augmented generation (RAG) pipelines, that means one small model can handle cross-modal retrieval, such as finding a video clip from a voice memo, without sending data to a server.

Key Features at a Glance

• Shared multimodal space: Code, images, video, and audio are embedded together, so one model handles cross-modal search.
• Modular sizing: Text-only workloads need as little as 270M parameters, with optional vision (170M) and audio (300M) encoders for full multimodal support.
• Storage reduction: Matryoshka Representation Learning lets developers truncate 768-dimension vectors to 512, 256, or 128, for up to 6x storage reduction in local vector databases.
• Low memory footprint: With quantization on a Google Pixel 11 Pro, it needs as little as ~191MB active RAM for text-only weights and ~567MB for the full multimodal model.
• Longer context: The 8K token context window is 4x that of EmbeddingGemma 1 and can process up to 5.5 minutes of audio, 29 images, or 58 video frames.
• Stronger code retrieval: MTEB Code scores rise 9.92 points, from 68.76 to 78.68, for local codebase indexing and coding agent retrieval.

Availability

EmbeddingGemma 2 launches today under the commercially permissive Apache 2.0 license, with model weights on Hugging Face and Kaggle, and Gemini Enterprise Agent Platform Model Garden availability is coming soon.