German search-infrastructure company Mixedbread has released Wholembed v3, a multimodal, multilingual retrieval model covering more than 100 languages and handling text, images, audio, and video. The company claims it's the first semantic model to surpass BM25 lexical retrieval on the LIMIT benchmark.
The secret sauce: a late-interaction multi-vector architecture, an idea popularized by ColBERT. Rather than compressing an entire document into a single vector, Wholembed v3 generates a separate low-dimensional vector for each token. That, Mixedbread argues, is where single-vector retrieval gets it wrong — it returns results that are sort of right, but not actually right.