NVIDIA Nemotron 3.5 Retrieval Suite Cuts Agent Inference Costs with Open Models

NVIDIA has expanded its Nemotron Retriever collection with five new open-weight models built for production agent pipelines: Nemotron 3.5 Embed in 8B and 1B BF16 variants plus a 1B NVFP4 quantized version, Nemotron 3.5 Rerank 8B BF16, and the lightweight Nemotron 3.5 Parse at 631M parameters. All carry a permissive commercial license, letting enterprises deploy them without royalty concerns.

Why retrieval quality drives infrastructure economics

Agentic workflows spend a disproportionate share of inference budget on repeated searches when initial retrieval misses key evidence. NVIDIA's benchmarks on ViDoRe v3 (multimodal), BRIGHT, and BrowseComp-Plus (text-only) show that adding Nemotron 3.5 Rerank on top of the 8B or 1B embedder cuts total agent tokens per task by roughly 13–14% while lifting nDCG@10 nearly a full point above Qwen3-VL-Reranker-8B at comparable token cost. For GPU clusters serving thousands of concurrent agents, that reduction translates directly into lower GPU-hour consumption and higher throughput.

Nemotron 3.5 Parse rounds out the stack by extracting structured content — tables, charts, diagrams — from PDFs and scanned documents, feeding richer evidence into the embed-and-rerank pipeline. The combined suite lets developers build retrieval-augmented agents that reason over heterogeneous enterprise data without chaining multiple proprietary services.