Granite Embedding Multilingual R2

"Granite Embedding Multilingual R2 Models." Two ModernBERT-based encoder embedding models — 311M (full) and 97M (compact) parameters — supporting 200+ languages and a 32K context window (64x the R1 generation). Reports SOTA on multilingual retrieval and code retrieval benchmarks.

Refresh of IBM Granite's embedding tier targeting enterprise multilingual retrieval. The 32K context jump from R1 makes the family viable as a long-document embedding backbone alongside the Granite 4.x generative line.