Mixedbread Wholembed v3

### TL;DR

Mixedbread Wholembed v3 is a unified omnimodal multilingual late-interaction retrieval model that sets a new state-of-the-art across text, audio, images, and video. It is the first semantic model to outperform lexical-based retrieval on LIMIT benchmark and powers Mixedbread Search as the default model.

Key Insights & Metrics

Key Features

• First semantic retrieval model to surpass lexical BM25 on the LIMIT benchmark
• Unified retrieval across text, audio, images, and video in 100+ languages
• 2M free tokens for new users; default model for all new Mixedbread Search stores

Discussion