Nvidia's Nemotron 3 Embed ranks first overall on a new retrieval benchmark

8:00 AM · July 16, 2026

Nvidia released Nemotron 3 Embed, a text embedding model designed to serve as a foundational component in retrieval augmented generation systems and agentic search workflows, according to a post on Hugging Face's blog. The model generates dense vector representations of multilingual text that can be used for retrieval, semantic search, and the kind of document lookup that underpins many AI agents, and Nvidia says it achieves state of the art performance among models of comparable size across multiple multilingual retrieval benchmarks. On the Retrieval Embedding Benchmark, a newer evaluation designed to more reliably measure how well embedding models perform in real world retrieval settings rather than on narrower academic tasks, Nemotron 3 Embed ranked first overall across sixteen public tasks. Nvidia has made the model's weights openly available on Hugging Face alongside details of its training approach, continuing the company's practice of releasing its Nemotron family as open models that developers can inspect and deploy directly rather than only access through a hosted API. The release is a comparatively small, technical one next to the frontier model announcements dominating the week's headlines, but embedding models like this one sit underneath a large share of the retrieval and search infrastructure that AI agents depend on to find relevant information before generating an answer.