EmbeddingGemma 2 Release Notes

Google DeepMind has launched EmbeddingGemma 2, a lightweight multimodal embedding model that maps text, images, audio, and video into a unified embedding space. This model is designed specifically for on-device inference, enabling privacy-first retrieval augmented generation (RAG) and cross-modal semantic search on consumer hardware.

Technical Architecture and Efficiency

EmbeddingGemma 2 is built on the Gemma 4 architecture and released under an Apache 2.0 license. It consists of 740 million parameters in its full configuration, but utilizes a modular design to optimize resource consumption based on the workload:

  • Modular Parameter Distribution: Text-only workloads require only 270M parameters. Full multimodal support is achieved by adding an optional vision encoder (170M parameters) and an audio encoder (300M parameters).
  • Storage Optimization via MRL: The model employs Matryoshka Representation Learning (MRL), allowing developers to truncate output vectors from 768 dimensions down to 512, 256, or 128. This can result in up to a 6x reduction in memory usage and local vector database storage.
  • On-Device Memory Footprint: When quantized and run on a Google Pixel 11 Pro, the model requires approximately 191MB of active RAM for text-only weights and 567MB for the full multimodal model.

Capabilities and Performance

EmbeddingGemma 2 provides a unified embedding space for multiple modalities, allowing users to perform tasks such as searching audio recordings via text queries or finding specific video clips from voice memos.

Benchmarks and Quality

EmbeddingGemma 2 achieves leading scores among sub-1B multimodal embedders. Key performance metrics include:

  • Code Performance: The model shows a 9.92-point improvement on the MTEB Code benchmark, increasing from 68.76 (EmbeddingGemma 1) to 78.68.
  • General Performance: It matches or outperforms many larger models across text, vision, and audio tasks, setting a new quality-per-parameter standard for models under 1B parameters.

Context Window

The model features an 8K token context window, which is four times larger than the first version of EmbeddingGemma. This expanded window allows the model to process the following directly on local hardware:

  • Up to 5.5 minutes of audio.
  • Up to 29 images.
  • Up to 58 video frames.
  • Interleaved combinations of the above modalities.

Integration and Deployment

Because EmbeddingGemma 2 shares its text tokenizer and audio encoder with Gemma 4, it can be integrated into a unified pipeline with Gemma 4 generative models to reduce the combined total memory footprint for on-device RAG.

Deployment Tools

EmbeddingGemma 2 is supported across several platforms and frameworks:

  • Model Weights: Available on Hugging Face and Kaggle.
  • On-Device Frameworks: Google AI Edge MediaPipe and LiteRT.
  • Web Deployment: transformers.js and WebGPU.
  • Serving Stacks: transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, and LMStudio.
  • Vector Storage: Qdrant.
  • Fine-tuning: Guidance is provided via Unsloth.

Sources