Gemini 4 Carbon: Our next frontier model
deepmind.google·5d ago
AI Engineering
Hugging Face Blog·May 14·5 min read
Granite Embedding Multilingual R2: Open Apache 2.0 Multilingual Embeddings with 32K Context — Best Sub-100M Retrieval Quality
TL;DR
Multilingual embedding models often require a trade-off between size and language coverage. Granite Embedding Multilingual R2 introduces two models that support over 200 languages with improved retrieval quality and a context length of 32,768 tokens.
✦ Why It Matters
Engineers can integrate advanced multilingual capabilities into their applications with minimal effort and improved performance.
Key Takeaways
- Use the 97M model for efficient multilingual retrieval tasks.
- Leverage 32K-token context for processing long documents.
- Integrate with existing frameworks using a simple model name change.
- Explore Matryoshka embeddings for flexible dimensionality in applications.
Full Summary
Granite Embedding Multilingual R2 features two models: a 311M full-size model and a 97M compact model, both built on the ModernBERT architecture. These models support over 200 languages and are specifically tuned for 52 languages, enhancing retrieval quality significantly.
The 97M model scores 60.3 on the MTEB Multilingual Retrieval benchmark, the highest for any open model under 100M parameters, while the 311M model scores 65.2, ranking second among models under 500M parameters. Key improvements include a 32K-token context capability, allowing for better handling of long documents.
Both models are designed for easy integration into existing frameworks like LangChain and Haystack, requiring minimal changes for deployment.