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

NVIDIA Releases Open Embedding Models, Claims Top Benchmark Spot

NVIDIA has released Nemotron 3 Embed, a new family of open, commercially available embedding models designed for enterprise retrieval-augmented generation (RAG) and agentic AI workflows. The flagship 8-billion parameter model, Nemotron-3-Embed-8B-BF16, has secured the #1 overall rank on the Retrieval Text Embedding Benchmark (RTEB), scoring 78.5%. The release addresses a critical bottleneck in multi-step AI agents, where poor retrieval quality can lead to irrelevant context, wasted token budgets, and flawed reasoning.

A Trio of Models for Quality and Efficiency

The Nemotron 3 Embed collection includes three distinct models to provide developers with a range of options balancing performance and deployment cost. The 8B model sets the quality standard, while two smaller 1.14B parameter variants cater to production environments where latency and throughput are primary concerns. The Nemotron-3-Embed-1B-NVFP4 model is specifically optimized for NVIDIA's Blackwell architecture, using 4-bit quantization to reduce its memory footprint and increase throughput while retaining over 99% of the BF16 model's accuracy.

• Open and Adaptable: The models are released with open weights and fine-tuning recipes, allowing teams to adapt them to proprietary data and deploy on their own infrastructure.
• Expanded Context: A 32,000-token context window enables retrieval over long documents and multi-turn agent histories without truncation.
• Ecosystem Integration: The models are available immediately through Hugging Face, as an NVIDIA NIM microservice, and with support from vLLM and other AI cloud partners.

Demonstrated Impact on Agentic Efficiency and Enterprise Adoption

NVIDIA's internal evaluations show a direct link between higher retrieval accuracy and lower downstream costs in agentic systems. By providing more relevant information upfront, the Nemotron 3 Embed models help AI agents complete tasks with fewer search queries and reasoning steps, thereby reducing overall token consumption. The models are already being evaluated by enterprise partners, including Automation Anywhere, Boomi, IBM, and Palantir. In one proof-of-concept, IBM reported that the 1B model outperformed competitors on the LongMemEval benchmark within its watsonx.data platform.

Strategic Takeaway: With Nemotron 3 Embed, NVIDIA further solidifies its full-stack enterprise strategy by delivering state-of-the-art open models that are finely tuned for its own Blackwell hardware and NIM serving infrastructure, creating a powerful, vertically integrated ecosystem that is difficult for competitors to replicate.