NVIDIA Nemotron 3 Embed Ranks #1 Overall on RTEB, Advancing Agentic Retrieval

Source: HuggingFace Date: 2026-07-16 URL: https://huggingface.co/blog/nvidia/nemotron-3-embed-wins-rteb

Summary

NVIDIA released Nemotron 3 Embed, an open embedding model family (8B-BF16, 1B-BF16, 1B-NVFP4) built for RAG and agentic retrieval. The 8B flagship scored 78.5% on RTEB and 75.5% on MMTEB Retrieval, topping the RTEB leaderboard, while the 1B variant cuts error rate 27% over its predecessor for cheaper deployment. Models support 32k context and multilingual retrieval, with an NVFP4-quantized variant tuned for Blackwell hardware.

Implications

• Agentic engineering: better retrieval means agents surface relevant evidence earlier, cutting redundant re-queries and token spend — a direct lever on agent cost and latency, not just answer quality.
• Local models: the 1B/NVFP4 tier is small enough for local or edge RAG stacks, extending the local-first embedding option set alongside existing open models.
• Signals continued vendor investment in retrieval as its own competitive layer, separate from the generation model itself.