Granite Embedding

Granite Embedding is IBM's family of open text-embedding models, released December 18, 2024 under Apache 2.0. It ships in five sizes: 30M and 125M English models (RoBERTa-style, 512-token context) and 107M and 278M multilingual models (XLM-RoBERTa-style covering 12 languages: EN, DE, ES, FR, IT, JA, KO, NL, PT, ZH, AR, CS), plus a 30M sparse variant. They use CLS pooling and report competitive MTEB retrieval scores for their size.

Trained on a mix of web title-body pairs, permissively licensed public pairs, IBM-internal technical data, and synthetic data, they are designed for RAG, semantic search, and similarity workloads. Compatible with the sentence-transformers library. A second-generation R2 lineup (311M multilingual, rerankers) followed in 2025.