Google DeepMind releases EmbeddingGemma 2 multimodal embedding model
Key Takeaway
EmbeddingGemma 2 is an open-source, 740M parameter multimodal model designed for efficient on-device text, image, audio, and video embeddings.
Summary
Google DeepMind has launched EmbeddingGemma 2, a lightweight multimodal embedding model released under the Apache 2.0 license. Built on the Gemma 4 architecture, the 740M parameter model maps text, code, images, audio, and video into a unified embedding space. It is specifically optimized for on-device inference, supporting privacy-first RAG pipelines and local semantic search.
Technical specifications of EmbeddingGemma 2
- Architecture: 740M parameters, built on Gemma 4, with modular encoders (270M for text-only, 170M for vision, 300M for audio).
- Context Window: 8K tokens, enabling processing of up to 5.5 minutes of audio or 58 video frames.
- Storage Efficiency: Uses Matryoshka Representation Learning (MRL) to allow vector truncation from 768 down to 128 dimensions.
- Memory Footprint: Requires approximately 191MB RAM for text-only weights and 567MB for the full multimodal model on quantized hardware.
- Performance: Significant improvements in code tasks (MTEB Code score increased from 68.76 to 78.68).
Why it matters
The Apache 2.0 release allows developers to build private, offline, and low-latency multimodal search and RAG applications directly on consumer hardware.
Original reference: deepmind.google →