granite-embedding 278m

IBM's high-efficiency 278M parameter model optimized for multilingual text embeddings and high-performance RAG workflows.

The granite-embedding 278m model delivers a lean, high-performance solution for dense vector retrieval across 38 languages. Built on a transformer architecture with a 512-token context window, it balances a compact 278-million parameter footprint with state-of-the-art accuracy on the MTEB benchmark. IBM engineered this model specifically for Retrieval-Augmented Generation (RAG) tasks, ensuring low-latency inference and precise semantic search capabilities for enterprise AI pipelines.

https://huggingface.co/ibm-granite/granite-embedding-278m-multilingual