NVIDIA Nemotron 3.5 Embed, Rerank, and Parse: Advancing Agents with World-Class Retrieval Models
huggingface ·
NVIDIA has released the Nemotron 3.5 Retriever collection, featuring new open models for embedding, reranking, and document parsing to enhance enterprise AI agents. The suite includes Nemotron 3.5 Embed in 8B and 1B variants, with the latter offering NVFP4 quantization for high-throughput inference on NVIDIA hardware. These models achieve state-of-the-art performance on multimodal retrieval benchmarks like ViDoRe v3, effectively searching across text, charts, and tables. Additionally, the Nemotron 3.5 Rerank model improves agent efficiency by reducing token usage by approximately 14% while increasing retrieval accuracy. For document processing, Nemotron 3.5 Parse extracts structured content from complex PDFs, ranking first on the ParseBench leaderboard among sub-billion parameter models. Early evaluations from partners such as Elastic, Vectorize, and Perplexity highlight significant gains in recall and relevance for long-context and domain-specific tasks. Developers can access these weights via Hugging Face or deploy them using NVIDIA NIM for production-grade serving.
huggingface