NVIDIA Adds Nemotron 3.5 Embed, Rerank and Parse to Its Retriever Lineup

The lineup comprises Nemotron 3.5 Embed in 8B BF16, 1B BF16 and 1B NVFP4 variants, Nemotron 3.5 Rerank 8B BF16, and Nemotron 3.5 Parse at 631 million parameters, according to a blog post published on Hugging Face.

NVIDIA frames retrieval quality as both an accuracy problem and a cost problem. When the evidence an agent retrieves is incomplete or irrelevant, the agent repeats searches and inspects additional documents, consuming more inference tokens before it completes the task. Reranking, the company says, improves the quality of the evidence that enters the agent's context.

The company illustrated the requirement with a stockout analysis: an agent asked to identify products at risk of running out next quarter must combine demand forecasts held in charts, inventory levels held in tables and supplier lead times held in reports, then reason about whether replenishment will arrive before inventory is depleted. Doing so requires locating the relevant evidence across documents and using it together.

To measure how retrieval affects that process, NVIDIA evaluated Nemotron 3.5 Embed and Rerank on ViDoRe v3, which uses combined text and image inputs, and on the text-only benchmarks BRIGHT and BrowseComp-Plus. In the tested pipeline, an agent queries a corpus through a search tool, inspects the returned evidence and refines its next query. The underlying LLM, tools, task set, corpus and execution policy were held fixed while only the retrieval configuration changed, covering embedding alone, embedding plus reranking, and different models for each role. The company measured task success and total agent inference tokens per task, counting input and output tokens across all turns.

NVIDIA reported that adding Nemotron 3.5 Rerank reduced agentic token cost per query by approximately 14% with Nemotron 3.5 Embed 8B and by 13% with Nemotron 3.5 Embed 1B, while retrieval accuracy improved in both configurations. With the 8B embedder as the retriever, the reranker delivered roughly 0.9 points higher nDCG@10 than Qwen3-VL-Reranker-8B at a similar agentic token cost, the company said.

The embedder extends the previous generation's Nemotron 3 Embed text models to retrieve evidence from document images, including tables, charts, infographics, datasheets and schematics, so applications can search visual content alongside document text. NVIDIA said text retrieval was evaluated across 34 languages, while multilingual visual retrieval was evaluated on MIRACL Vision, covering 18 languages, and on KoViDoRe; the models may support additional languages beyond those tested.

The variants are aimed at different deployment settings. The 8B embedder targets quality-focused search over text and visually rich documents; the 1B BF16 model distills the 8B model's multimodal retrieval into a smaller footprint for indexing and query serving; the 1B NVFP4 variant applies four-bit quantization to both text and image encoders for large-scale document indexing and high-throughput retrieval on supported hardware. Nemotron 3.5 Rerank 8B is a cross-encoder that jointly evaluates query-document pairs to refine relevance ranking across text and image candidates before evidence is passed to retrieval-augmented generation and agent workflows.

The announcement did not disclose pricing, availability dates or licensing terms beyond describing the models as open and commercially available.