TL;DR

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

NVIDIA has expanded its Nemotron Retriever collection with open models for multimodal embedding, reranking and document parsing. The company reports improved retrieval results and lower agent inference-token use in specific evaluations, but the supplied announcement does not provide full details for the Parse model or independent validation of the benchmarks.

NVIDIA has expanded its Nemotron Retriever collection with five open models for embedding, reranking and document parsing, aiming to help enterprise agents find useful evidence across text and image-rich documents. The company says its retrieval configurations improved benchmark performance while reducing agent inference-token use in selected tests; those results are company-reported and depend on the stated evaluation setup.

The new collection includes Nemotron 3.5 Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16 and Parse at 631 million parameters. NVIDIA describes the embedding models as retrieving information from text and visual material such as tables, charts, infographics, datasheets and schematics. The reranker scores query-document pairs to help prioritize relevant results for an agent’s context.

In tests on ViDoRe v3, BRIGHT and BrowseComp-Plus, NVIDIA compared retrieval setups while holding the language model, tools, tasks, corpus and execution policy fixed. The company reports that adding Nemotron 3.5 Rerank cut agent inference tokens per query by about 14% with Embed 8B and 13% with Embed 1B, while also improving retrieval accuracy. With Embed 8B, it says its reranker scored about 0.9 points higher in nDCG@10 than Qwen3-VL-Reranker-8B at a similar agent token cost.

On ViDoRe v3.1, NVIDIA reports an nDCG@10 score of 61.17 for Embed 8B, compared with 52.23 for Llama Nemotron Embed VL 1B. Its Embed 1B BF16 model scored 58.38, which the company describes as retaining 95.4% of the 8B model’s score. These are reported benchmark outcomes, not a guarantee of performance on other workloads.

Retrieval Costs Shape Agent Work

Agents that repeatedly search or inspect irrelevant material can use more inference tokens before reaching an answer. For companies running agents over large document collections, better retrieval could reduce that repeated work while giving the model evidence that is more relevant to the task. NVIDIA’s reported results connect retrieval quality with both task success and token use, rather than treating search as a separate preprocessing step.

The multimodal focus also addresses a practical limitation in document-heavy workflows: important details may be in a chart or table image rather than in searchable text. A system that can retrieve such content could help agents answer questions that require evidence from multiple formats. The real-world effect will depend on the document collection, hardware, workload and integration choices.

From Text Search to Visual Evidence

The announcement presents Nemotron 3.5 Embed as an extension of the prior Nemotron 3 Embed text embedding models. Embedding models represent queries and documents so search systems can find semantically related material; rerankers then compare candidate results against a query to refine their order. NVIDIA’s stated workflow uses an agent that searches, inspects evidence and may revise its search based on what it finds.

The models offer different deployment options. The 8B BF16 embedding model is positioned for quality-focused retrieval, while the 1B BF16 version is a smaller multimodal option. The 1B NVFP4 version applies quantization to text and image encoders for inference efficiency on supported hardware. NVIDIA also says its embedding fine-tuning recipes can adapt models to custom collections; in chart and infographic experiments, it reports relative nDCG@10 gains of 18% for Embed 1B BF16 and 3% for Embed 8B BF16.

NVIDIA says text retrieval was evaluated across 34 languages, while multilingual visual retrieval was tested on MIRACL Vision across 18 languages and on KoViDoRe. The source says support may extend beyond the languages evaluated, but does not establish results for those additional languages.

“Adding Nemotron 3.5 Rerank reduces agentic token cost per query by approximately 14% with Nemotron 3.5 Embed 8B, and 13% with Nemotron 3.5 Embed 1B, while improving retrieval accuracy in both configurations.”

— NVIDIA, in the Hugging Face announcement

Benchmark Scope and Parse Details

The supplied announcement does not specify the publication date, provide full benchmark methodology or results tables, or include independent replication of the reported scores and token savings. It also names Nemotron 3.5 Parse and its parameter count but gives little detail about its capabilities, supported formats, availability or evaluation. Pricing, licensing terms, hardware requirements beyond the NVFP4 deployment note, and the timing of broader availability are not stated in the source material.

Benchmark performance may differ on other corpora, languages and agent tasks. The announcement says additional languages may be supported, but reports evaluations only for the stated language sets.

Availability and Independent Testing

The announcement describes the models as open and commercially available, but the supplied source does not give release links, licensing details or a schedule for further updates. The next useful checks are the model documentation and artifacts, where users can examine deployment requirements and reproduce the reported evaluations.

Independent testing on enterprise document collections would help establish whether the reported retrieval gains and token reductions carry over to different workloads. More information about Nemotron 3.5 Parse would also clarify how it fits alongside the embedding and reranking models.

Where I land

My assessment is that this release targets a real bottleneck in agent systems: retrieval affects both the evidence an agent sees and the amount of work it may spend searching. The reported token reductions and multimodal benchmark scores make the models worth evaluating for document-heavy workflows, especially where relevant information appears in charts and tables.

The strongest counterargument is that benchmark gains may not translate to a company’s own documents, and the supplied material leaves important questions about methodology, deployment and the Parse model unanswered. I would put more weight on the efficiency claims after independent comparisons on varied enterprise collections show similar accuracy and token savings, with clear licensing and hardware requirements.

Key Questions

Which Nemotron 3.5 models were announced?

NVIDIA listed three embedding variants—8B BF16, 1B BF16 and 1B NVFP4—plus Rerank 8B BF16 and Parse at 631 million parameters.

What did NVIDIA report about inference-token use?

In its evaluated retrieval setups, NVIDIA says adding Nemotron 3.5 Rerank reduced agent inference tokens per query by about 14% with Embed 8B and 13% with Embed 1B, while improving retrieval accuracy.

Can the embedding models search charts and tables?

NVIDIA says the embedding models retrieve evidence from text and images, including tables, charts, infographics, datasheets and schematics. Performance will depend on the documents and application.

Are the reported benchmark results independently verified?

The supplied announcement presents them as NVIDIA’s evaluations. It does not provide independent validation or enough methodology detail to establish how results transfer to other workloads.

Source: Hugging Face