XNews

NVIDIA Releases Nemotron 3.5 Embed, Rerank, and Parse Retrieval Models

NVIDIA Releases Nemotron 3.5 Embed, Rerank, and Parse Retrieval Models

NVIDIA released the Nemotron 3.5 Retriever collection, including Embed 8B BF16, 1B BF16, 1B NVFP4, Rerank 8B BF16, and Parse at 631M parameters, as open, commercially available models for multimodal retrieval, relevance ranking, and document processing in enterprise agent workflows. The company stated that better retrieval helps agents reach grounded answers with fewer wasted inference tokens, and it evaluated Embed and Rerank across ViDoRe v3, BRIGHT, and BrowseComp-Plus.

Event version 1.

Key insights

New model lineup

The collection includes Nemotron 3.5 Embed 8B BF16, 1B BF16, and 1B NVFP4, Nemotron 3.5 Rerank 8B BF16, and Nemotron 3.5 Parse at 631M parameters.

Token efficiency rationale

NVIDIA stated that incomplete or irrelevant evidence causes agents to repeat searches and inspect more documents, consuming more inference tokens, and that better retrieval can reduce this waste.

Evaluation benchmarks

Nemotron 3.5 Embed and Rerank were evaluated across ViDoRe v3, BRIGHT, and BrowseComp-Plus, with ViDoRe v3 using combined text and image inputs and the other two using text-only inputs.

Multimodal retrieval focus

The models bring multimodal retrieval, relevance ranking, and document processing to enterprise agent workflows, with Nemotron 3.5 Embed and Rerank leading multimodal dense retrieval on ViDoRe v3.

Source posts