According to CoinMeta, Mixedbread has released a full-modal, multi-language retrieval model called Wholembed v3, which becomes the first semantic model to surpass BM25's lexical retrieval on the LIMIT benchmark. This model adopts a late-interaction multi-vector architecture and supports text, image, audio, and video retrieval, covering over 100 languages. Mixedbread has also released a dedicated search tool called agent toast 1, aimed at large models such as GPT and Claude. The official claim is that its search quality is close to that of cutting-edge proprietary models, but it offers faster speeds and lower costs. toast 1 has been made open-source through API, and the accompanying search framework is also open-source, although the model weights are not yet available for public access.