NVIDIA Nemotron 3 Embed Ranks #1 Overall on RTEB, Advancing Agentic Retrieval
18:01 · July 16, 2026 · Hugging Face Blog
Summary
NVIDIA has released the Nemotron 3 Embed collection of open embedding models, which includes an 8B-parameter flagship and smaller 1B variants optimized for production retrieval workloads. The 8B model, released in BF16 precision, currently ranks first on the RTEB leaderboard and delivers the highest average NDCG@10 across RTEB, ViDoRe V3 Text, MMTEB Retrieval, and LongEmbed benchmarks. The collection targets retrieval-augmented generation, agentic retrieval, code retrieval, and agent memory use cases, with explicit attention to both retrieval quality and downstream inference cost.
The 8B model starts from the Ministral-3-8B-Instruct-2512 backbone, which is converted from a causal decoder into a bidirectional encoder. It undergoes contrastive pre-training on web-sourced and synthetic text pairs, followed by fine-tuning on curated multilingual datasets spanning legal, finance, medical, business, and education domains. The 1B models are produced through a two-stage compression pipeline: a 3B bidirectional retriever is first derived from Ministral-3-3B-Instruct-2512, then successively pruned with NVIDIA ModelOpt’s mcore_minitron Neural Architecture Search and distilled from 8B teacher checkpoints using combined cosine-distance and mean-squared-error losses. The final 1.14B model employs Quantization-Aware Distillation and native NVFP4 quantization of linear-layer weights and activations to support efficient inference on Blackwell hardware.
In agentic settings, the models are evaluated with a search agent powered by Nemotron 3 Ultra. Stronger retrieval reduces estimated downstream token cost per query on ViDoRe V3, BRIGHT, and BrowseComp-Plus by returning relevant evidence earlier and thereby limiting repeated searches and reasoning turns. For high-throughput serving, NVIDIA provides an optimized NIM microservice for the 1B model that matches or exceeds vLLM throughput on GB200 and RTX PRO 6000 GPUs at input-sequence lengths of 256 and 1024 tokens. Open-source NeMo AutoModel fine-tuning recipes are also released; on the NV Docs evaluation set, fine-tuning the 1B BF16 model raises NDCG@10 from 56.7 % to 63.3 % and Recall@5 from 56.1 % to 62.8 %.
Why it matters
Provides concrete architectural, training, quantization, and serving details plus quantitative benchmarks directly usable by ML engineers building production retrieval, RAG, and agentic systems.
