NVIDIA's Nemotron 3 Embed tops the RTEB leaderboard

NVIDIA has released Nemotron 3 Embed 1B, a text embedding model that the company says ranks first overall on the Retrieval Embedding Benchmark (RTEB). At roughly one billion parameters, it is small enough to run economically while aiming for state-of-the-art retrieval quality, and it ships in BF16 on Hugging Face.

Embedding models are the quiet workhorses of modern AI systems. They convert text into dense vectors that power semantic search, recommendation, and the retrieval step in retrieval-augmented generation (RAG). A better embedding model directly improves how relevant the documents a system pulls in are, which in turn shapes the quality of everything downstream.

Why it matters

RTEB is designed to measure retrieval performance in ways meant to reflect practical deployment rather than benchmark overfitting, so a top ranking there is a meaningful signal for teams building search and RAG pipelines. A few things stand out about this release:

• It leads the RTEB leaderboard overall despite a modest 1B parameter footprint.
• Its small size makes it cheaper to serve at scale than many larger embedding models.
• NVIDIA is releasing open weights, giving developers a strong self-hostable option.

The model is distributed under a custom NVIDIA license rather than a standard open-source license, so teams should review the terms before production use. Full details and evaluation results are in NVIDIA's announcement post.