NVIDIA Nemotron 3.5 Embed, Rerank, And Parse: Advancing Agents With World-Class Retrieval Models

🔍 Read the full analysis: NVIDIA Nemotron 3.5 Embed, Rerank, And Parse: Advancing Agents With World-Class Retrieval Models on ThorstenMeyerAI.com

NVIDIA has expanded its Nemotron Retriever collection with five open models for embedding, reranking and document parsing, targeting enterprise agents that search text- and image-rich documents. As detailed in the original analysis, the company says selected evaluations showed better retrieval accuracy and lower agent inference-token use, though the figures have not been independently validated in the supplied material.

The five models are 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 relevant material from text as well as visual content in tables, charts, infographics, datasheets and schematics. The reranker scores query-document pairs to help an agent prioritize candidate evidence.

NVIDIA says its evaluations used ViDoRe v3, BRIGHT and BrowseComp-Plus, holding the language model, tools, tasks, corpus and execution policy fixed when comparing retrieval configurations. The company reports that 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 in both configurations.

For another comparison, NVIDIA reports an nDCG@10 score of 61.17 for Embed 8B on ViDoRe v3.1, against 52.23 for Llama Nemotron Embed VL 1B. Embed 1B BF16 scored 58.38, which NVIDIA describes as 95.4% of the 8B model’s score. These are company-reported results tied to the stated tests, not evidence that the same performance will hold across other collections or workloads.

Retrieval Gains Could Cut Agent Work

Retrieval quality can affect both the evidence an agent uses and the amount of model processing needed to find it. If an agent repeatedly retrieves irrelevant files or misses useful passages, it may spend more inference tokens searching before it can respond. NVIDIA’s reported results connect retrieval accuracy with token use, making the release relevant to organizations running agents over large document collections.

The multimodal emphasis addresses a practical gap in document search: important information may be embedded in a chart or table image rather than represented in searchable text. Models that retrieve such material could help agents assemble evidence across different formats. The impact for a particular business will depend on its documents, hardware, workloads and integration; benchmark results alone do not establish savings or improved outcomes in production.

From Semantic Search to Visual Documents

Embedding models represent queries and documents in a form that lets a search system find semantically related material. A reranker then compares a query with candidate results and can change their order before the agent uses them. In NVIDIA’s described workflow, an agent searches, inspects evidence and may revise its search based on what it finds. The new collection extends the company’s earlier Nemotron 3 Embed text models toward multimodal retrieval.

NVIDIA positions the 8B BF16 embedding model for quality-focused retrieval and the 1B BF16 model as a smaller multimodal option. The 1B NVFP4 variant uses quantization for its text and image encoders, which the company presents as an inference-efficiency choice 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.

The company says it evaluated text retrieval across 34 languages, and multilingual visual retrieval on MIRACL Vision across 18 languages and on KoViDoRe. The supplied information does not establish performance for languages beyond those evaluated. It also gives little background or technical detail about Parse beyond naming the model and its parameter count.

“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 its Hugging Face announcement

Benchmark Limits and Parse Details

The supplied announcement does not give a publication date, full benchmark methodology or complete results tables, and it includes no independent replication of the reported scores or token savings. The stated comparisons describe particular evaluation setups; performance on other corpora, languages and agent tasks remains uncertain.

Details about Nemotron 3.5 Parse are especially limited: the material provides its parameter count but not its capabilities, supported document formats, evaluation results or how it works alongside the embedding and reranking models. The supplied source also does not give pricing, licensing terms, complete hardware requirements or release links. Although NVIDIA describes the models as open and commercially available, those practical details are not specified here.

Documentation and Independent Tests

Users and developers will need the model artifacts and documentation to check availability, licensing, supported hardware and deployment requirements. Independent tests on enterprise document collections could show whether NVIDIA’s reported retrieval improvements and token reductions carry over to different settings.

Further technical information about Parse would clarify its role in the collection, including which inputs it handles and how its output connects to retrieval workflows. The supplied announcement does not state a schedule for additional documentation or updates, so no next release milestone is confirmed.

Key Questions

What did NVIDIA announce?

NVIDIA expanded its Nemotron Retriever collection with five models: three embedding variants, one reranker and Nemotron 3.5 Parse, listed at 631 million parameters.

What does the reranker do?

A reranker scores query-document pairs to help order candidate search results by relevance before an agent uses them. NVIDIA says its tests found improved retrieval accuracy and lower agent inference-token use with the reranker added.

How much lower was token use in NVIDIA’s tests?

NVIDIA reports about 14% fewer agent inference tokens per query with Embed 8B and about 13% fewer with Embed 1B when adding Nemotron 3.5 Rerank. These are company-reported results from selected evaluations, not a guarantee for other workloads.

What is known about Nemotron 3.5 Parse?

The supplied announcement identifies Parse as a 631-million-parameter model but does not provide full details about its capabilities, supported formats, availability or evaluation.

Have the benchmark results been independently confirmed?

Not in the supplied material. NVIDIA reports the benchmark scores and token reductions, but the source provides no independent replication or complete methodology and results tables.

Primary source: Hugging Face · via ThorstenMeyerAI.com