NVIDIA Nemotron 3 Embed 8B takes the top RTEB retrieval slot

Retrieval becomes an agent cost control point

In RAG systems and search-based agents, the embedding model often decides whether the answer is grounded before the language model begins writing. NVIDIA AI wrote on X that “Nemotron 3 Embed 8B” reached “#1 overall on RTEB.” The practical claim is that better retrieval gives agents more relevant context earlier, reducing repeated searches, extra reasoning turns, and wasted context inspection.

The concrete result is 78.5%. NVIDIA's Hugging Face post says Nemotron-3-Embed-8B-BF16 scores 78.5% on RTEB and 75.5% on MMTEB Retrieval. It also says the 1B BF16 variant scores 72.4% on RTEB, cutting error rate by 27% compared with its prior 1B predecessor, and scores 71.0% on MMTEB Retrieval, reducing error rate by 28%. That gives teams a quality ceiling and a smaller production option in the same family.

The release is packaged for deployment rather than only leaderboard attention. NVIDIA describes an 8B flagship model, a 1B BF16 model for latency- and cost-sensitive retrieval, and a 1B NVFP4 model optimized for Blackwell high-throughput serving with a smaller memory footprint. The post also lists open weights, datasets, recipes, a 32k context window, multilingual and code retrieval, fine-tuning and distillation recipes, NVIDIA NIM availability, vLLM support, and cloud inference partner access.

NVIDIA AI's account has increasingly framed retrieval as part of agent infrastructure, not a narrow search feature. The next thing to watch is transfer from RTEB into messy enterprise corpora: internal documents, code repositories, ticket histories, and long multi-turn agent logs. The 1B models may matter as much as the 8B leader if they lower serving cost enough to make retrieval upgrades practical. The source tweet and technical post provide the result details.