March 2026 Summaries

Mixedbread Search v3 is presented as a retrieval system intended to reduce the “gap to oracle,” or the performance difference between systems using retrieved evidence and those given the correct documents directly, across agentic research and enterprise workflows. On BrowseComp-Plus, a multi-hop web research benchmark, it reportedly ranks first under both standard and document-access retrieval scaffolds, achieving 90.48% accuracy in the stronger setting versus a 93.5% oracle score. On MADQA, which evaluates question answering over heterogeneous and multimodal PDF collections, Mixedbread-supported systems reportedly lead relevant leaderboard categories, with a Gemini 3 Pro one-shot system reaching 88.2% accuracy and a Distyl AI system called Button reaching 91.7%. In OfficeQA-Pro, a financial-document benchmark tested with OpenAI Codex, Mixedbread retrieval reached 64.42% correctness, within 0.99 points of the oracle configuration, while reducing latency and tool calls relative to a corpus-based baseline. The company notes that retrieval remains only one source of failure in difficult knowledge-work tasks, alongside ambiguity, missing evidence, and reasoning limitations, and offers its search service through an API designed to manage multimodal ingestion, indexing, and retrieval without requiring users to configure chunking, embeddings, vector databases, or reranking.

Mixedbread has released Wholembed v3, a unified late-interaction retrieval model designed for multilingual, multimodal search across text, images, audio, and video, which will power all new Mixedbread Search stores by default. The company positions the model as a response to limitations in traditional semantic retrieval, particularly for complex agentic applications and real-world data such as OCR documents, scanned PDFs, screenshots, and instructional videos. Wholembed v3 reportedly achieved state-of-the-art results on the LIMIT benchmark, where it surpassed BM25 lexical retrieval with 92.45 Recall@5 versus 85.7, and was also evaluated for supporting deep-research agents on BrowseComp-Plus. Mixedbread attributes its performance to precise candidate discrimination and robustness to noisy, heterogeneous information, while claiming partner testing showed it could replace more complex retrieval pipelines across several domains. The model is publicly available through Mixedbread Search, includes built-in audio and video support, offers new users two million free tokens, and is accessible through the platform or API.