2026 Climate Tech Companies to Watch
technologyreview.com·6d ago
AI Engineering
DeepMind Blog·6d ago·5 min read
EmbeddingGemma 2: an open, lightweight multimodal embedding model
TL;DR
EmbeddingGemma 2 is a new multimodal embedding model that integrates text, code, images, video, and audio into a single framework, achieving high performance with 740 million parameters. It enables advanced on-device search and retrieval capabilities, significantly enhancing user applications.
✦ Why It Matters
Start integrating EmbeddingGemma 2 into your applications for enhanced multimodal search capabilities today.
Key Takeaways
- Download EmbeddingGemma 2 from Hugging Face or Kaggle.
- Utilize Google AI Edge MediaPipe for on-device deployment.
- Fine-tune the model using Unsloth's guidance.
- Explore cross-modal search capabilities for your applications.
Full Summary
EmbeddingGemma 2 builds on the success of its predecessor, offering a lightweight solution for generating high-quality embeddings across multiple data types, including text, code, images, video, and audio. With 740 million parameters, it is optimized for on-device inference, allowing applications to process complex queries locally.
The model achieves a significant 9.92-point improvement in code performance, making it ideal for tasks like semantic code search. It also features a modular design, enabling developers to use as few as 270 million parameters for text-only tasks.
Storage efficiency is enhanced through Matryoshka Representation Learning, allowing dynamic truncation of output vectors. With an 8K token context window, it can handle extensive data inputs, making it suitable for various applications while ensuring data privacy and reducing latency.