Bekko Embedding is a family of compact multilingual retrieval models built through layer pruning and contrastive learning. Its smaller variant, bekko-a8m, scores 56.2 on MMTEB’s 18 retrieval tasks with only 7.67 million active parameters, outperforming larger models such as multilingual-e5 while processing 364 documents per second on CPU. Both variants support more than 100 languages, produce 384-dimensional embeddings that can be truncated further, and compress to 124-190 MiB, making browser-based inference without server uploads feasible.
