EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hugging Face Blog·Sep 3·5 min read
NeoMME: an efficient Multimodal-native and Multilingual Encoder
TL;DR
NeoMME is a new family of multilingual multimodal encoders that efficiently processes text and images using a single bidirectional Transformer. It achieves high performance in visual document retrieval while significantly reducing storage requirements.
✦ Why It Matters
Explore NeoMME for efficient visual document retrieval in your applications today.
Key Takeaways
- Implement NeoMME for faster visual document retrieval.
- Utilize late-interaction embeddings for improved query accuracy.
- Leverage hierarchical token pooling to reduce storage costs.
- Access NeoMME models on Hugging Face for immediate use.
How It Works
NeoMME processes text and images through a single Transformer encoder, allowing for efficient training and inference. It uses a masked discrete-diffusion objective for pretraining, where text tokens are masked to encourage the model to learn context from both text and images.
The model's architecture supports dynamic image resolution and long bidirectional context, enhancing its ability to handle complex documents.