Strengthening the retrieval layer for agents
NVIDIA announced three models aimed at agent-based AI applications: Nemotron 3.5 Embed, Rerank and Parse. The announcement was published on the Hugging Face blog on October 21, 2026.
Three models, three jobs
- Embed: Turns text into numerical vectors that support similarity search, forming the basis for finding relevant passages in large document collections.
- Rerank: Reorders the results returned by an initial search according to the query, keeping the most relevant content on top.
- Parse: Converts messy content in documents and files into structured text that a model can process.
Why it matters
AI agents frequently need to pull information from external sources while making decisions. When the retrieval step is weak, the agent acts on wrong or incomplete information. By offering the three models together, NVIDIA aims to let agent teams build their search pipeline under one roof.
Limited information
The source does not include parameter counts, licensing, benchmark results or access terms for the models. Those details are expected in the full announcement.
