News Hub

Mixedbread releases listwise v3.1 reranker matching GPT-5.6-sol quality at 61x lower latency.

Asymmetric quantization reduces late interaction retrieval corpus storage 32x by storing document vectors as binary signs.

vLLM runs on NVIDIA DGX Spark with unified memory, NVFP4 serving, and Prometheus metrics.

Mixedbread released mxbai-rerank-v3-listwise, a listwise reranker achieving state-of-the-art instruction following on all tested benchmarks.

Mixedbread Search v3 improves retrieval accuracy on BrowseComp-Plus, MADQA, and OfficeQA-Pro benchmarks.

Mixedbread released Wholembed v3, a unified multimodal multilingual retrieval model achieving state-of-the-art search performance across languages and modalities.

Mixedbread built multimodal late-interaction search achieving 80ms latency on billion-scale document collections.

Mixedbread released mxbai-edge-colbert-v0, a 17 million parameter ColBERT model outperforming ColBERTv2.

Mixedbread released a search API designed for AI with multi-modal and multi-lingual capabilities.

Mixedbread outlined their research vision and approach to solving search problems.

OCR errors limit RAG performance; multimodal vector stores outperform text-based approaches by 12%.

Mixedbread released mxbai-rerank-v2 with reinforcement learning and multilingual support.

Mixedbread releases mxbai-embed-xsmall-v1, a compact English embedding model with binary quantization support.

Mixedbread released Baguetter, an open-source framework combining sparse, dense, and hybrid retrieval methods.

Mixedbread introduced BMX, an improved BM25 search algorithm using entropy-weighted similarity and query augmentation.

Deepset and Mixedbread release deepset-mxbai-embed-large-v1, an open-source German/English embedding model.

Binary MRL achieves 64x embedding efficiency gain while retaining over 90% performance.

mxbai-colbert-large-v1 is a ColBERT model achieving state-of-the-art performance on 13 BEIR benchmarks for reranking and retrieval tasks.

Mixedbread released an English embedding model outperforming OpenAI's text-embedding-v3.

Mixedbread introduced 2D-Matryoshka embedding model reducing layers and dimensions while maintaining performance.