Granite Embedding Multilingual R2: Open Apache 2.0 Multilingual Embeddings with 32K Context
AI Summary
IBM has released two open-source, Apache 2.0-licensed multilingual embedding models built on the ModernBERT architecture: granite-embedding-311m-multilingual-r2 and granite-embedding-97m-multilingual-r2. Supporting over 200 languages, 52 explicitly tuned languages, and 9 programming languages, the models extend context windows to 32,768 tokens. The 97M compact model leads open sub-100M embedders with a 60.3 MTEB Multilingual Retrieval score, while the 311M model reaches 65.2 and supports Matryoshka dimensionality reduction.
Key Takeaways
- The 97M model outperforms existing open sub-100M multilingual embedders on MTEB Multilingual Retrieval.
- Both models natively support a 32K-token context window using ModernBERT encoder foundations and rotary position embeddings.
- Matryoshka representation learning allows the 311M model to scale down to 128 or 256 dimensions with minimal accuracy loss for faster vector searches.
Why it matters
These models solve the traditional trade-off between model size and multilingual capability, allowing developers to deploy high-throughput, enterprise-ready embeddings for cross-lingual search and RAG systems without restrictive non-commercial licenses.
Conversation 0
Be the first to share your perspective.