Granite Multilingual Embedding R2 Released

IBM Granite has released 97M and 311M multilingual embedding models R2. Both models are under the Apache 2.0 license, support 200+ languages and 9 programming languages for code search, and handle a 32,768 tokens context.

• granite-embedding-97m-multilingual-r2 has 97M parameters and 384-dim embeddings, achieving MTEB Multilingual Retrieval 60.3.
• granite-embedding-311m-multilingual-r2 has 311M parameters and 768-dim embeddings, achieving 65.2, and supports Matryoshka dimension reduction to 512/384/256/128.
• Retrieval/cross-lingual training was added for 52 languages and code, and it can be dropped in by just changing the model name in sentence-transformers, transformers, LangChain, LlamaIndex, Haystack, and Milvus.
• The 311M model is based on a 22-layer ModernBERT and was trained with knowledge distillation, contrastive fine-tuning, checkpoint merging, and Matryoshka, while the 97M model was made lightweight via vocabulary pruning and knowledge distillation.
• Training data mixed IBM-curated data, public web data, and synthetic data, while MS-MARCO and non-commercially licensed data were excluded.
• ONNX and OpenVINO weights for CPU inference are also provided.