The Pulse
NVIDIA Releases Five Nemotron 3.5 Retrieval Models
NVIDIA released five Nemotron 3.5 models for searching text and images, reranking results and extracting structure from documents. The company says the models cut agent search costs and lead selected retrieval benchmarks.
AI.info Team · Written with AI · fact-checked ·
NVIDIA announced five Nemotron 3.5 retrieval models on October 21, adding tools to search across text and images, rank candidate results and turn document pages into structured data. The release targets a specific bottleneck in AI agents: when search returns weak evidence, an agent may need more queries and more model tokens to finish a task. NVIDIA says the new models are available with open weights and can be used commercially.
Five models divide retrieval into three jobs
The lineup includes Nemotron 3.5 Embed 8B BF16, two 1B Embed versions in BF16 and NVFP4, Nemotron 3.5 Rerank 8B BF16, and Nemotron 3.5 Parse, a 631-million-parameter document-processing model. Embed searches text and visual material such as tables, charts, infographics, datasheets and schematics. Rerank scores the candidates returned by search against a query, while Parse extracts text and document structure for indexing and later retrieval.
NVIDIA says the embedding models were evaluated for text retrieval in 34 languages, and for visual retrieval in 18 languages on MIRACL Vision. The 8B reranker also handles text and image inputs, allowing a search system to rank a page image alongside text passages. The smaller Embed model offers a lower-footprint option; its NVFP4 version quantizes both the text and image encoders.
“Together, these state-of-the-art models bring multimodal retrieval, relevance ranking, and document processing to enterprise agent workflows.”
Yauhen Babakhin, Ryan Chesler, Jiarui Cai, Sean Sodha and Isabel Hulseman, NVIDIA
NVIDIA reports gains on retrieval benchmarks
On ViDoRe v3.1, NVIDIA reports that Embed 8B scored 61.17 nDCG@10, ranking first among dense retrieval models on the leaderboard as of October 20. The 1B BF16 model scored 58.38, which the company says is 95.4% of the 8B model’s score. For combined image-and-text inputs, Rerank 8B scored 60.49 nDCG@10 on ViDoRe v3.1, compared with 59.17 for Qwen3-VL-Reranker-8B.
NVIDIA also tested the models in an agent workflow where the language model, task set, corpus and execution policy stayed fixed while the retrieval configuration changed. Adding Rerank cut agent inference-token costs per query by about 14% with Embed 8B and 13% with Embed 1B, while improving retrieval accuracy in both configurations. With Embed 8B, NVIDIA says its reranker achieved about 0.9 points higher nDCG@10 than the Qwen reranker at a similar token cost.
NVFP4 and NIM target serving costs
On an NVIDIA GB200, Embed 1B NVFP4 processed 128.8 documents per second, up from 101.3 for the BF16 version, while retaining 99.85% of its retrieval accuracy on ViDoRe v3 with combined image and text. The figures describe NVIDIA’s test, not a general performance guarantee across hardware or workloads. The release gives teams a choice between a smaller quantized model and a higher-precision alternative.
NVIDIA also compared its NIM serving stack with a vLLM configuration on an RTX PRO 6000 Blackwell Server Edition. The company reported 6,031 query requests per second for NIM against 2,099 for vLLM, while describing passage-processing throughput as comparable. That result concerns the serving setup; it is separate from the retrieval-quality scores.
Parse brings scanned pages into the search pipeline
Parse is designed for PDFs and images where a text layer may omit information, disrupt reading order or separate table values from their headings. The model identifies document elements and their locations, then returns structured content such as Markdown, HTML tables and other blocks. NVIDIA reports a ParseBench score of 75.20 and a processing rate of 10.5 pages per second on one H100; it ranked first among models under one billion parameters as of October 20.
Developers can obtain the models, weights, data and training recipes through Hugging Face, or deploy them through NVIDIA NIM and listed inference partners. NVIDIA’s benchmark results and partner evaluations are company-reported; the announcement does not provide independent verification of those comparisons. The five releases span the retrieval pipeline from a scanned page to a ranked set of evidence ready for an agent.