EmbeddingGemma 2: an open, lightweight multimodal embedding model

Summary

Google DeepMind released EmbeddingGemma 2, an open-weight embedding model that maps text, code, images, video, and audio into one vector space. It is small enough to run on a phone and is licensed under Apache 2.0. The design is modular: the text-only part can be used on its own at a fraction of the full size.

Key points

• 740M parameters for the full multimodal model; the text-only path is 270M.
• 768-dimension vectors that can be truncated to 512, 256, or 128 (Matryoshka Representation Learning) to save storage.
• 8K-token input window, equal to about 5.5 minutes of audio, 29 images, or 58 video frames.
• On-device memory with quantisation: about 191 MB (text only) and 567 MB (multimodal), measured by Google on a Pixel 11 Pro.
• Google reports 78.68 on MTEB Code, 9.92 points above EmbeddingGemma 1, and leading MAEB (audio) and MIEB (image) scores among multimodal embedders under 1B parameters.
• Weights on Hugging Face and Kaggle; LiteRT builds for on-device use. Licence is Apache 2.0; access through Google's Gemini Enterprise Agent Platform is "coming soon".

Why it matters

A single small, permissively licensed model for cross-modal search makes on-device and self-hosted multimodal RAG practical without calling a cloud embedding API.