BAAI bge-reranker-v2.5-gemma2-lightweight is a roughly 9B multilingual reranker on Gemma 2 9B with token compression and layer cutoffs.
BGE-M3 is BAAI's open multilingual embedding model: MIT license, 1024-dim vectors, 8192-token inputs, and dense, sparse and multi-vector retrieval.
BAAI's AREX-2 is a 27B open-weight Apache-2.0 agent model that improves solutions over propose, measure, reflect, and revise rounds, with 262K context.