Nemotron 3 Embed Hits #1 on RTEB With NVFP4 Quantization

NVIDIA released the Nemotron 3 Embed collection, a suite of text embedding models engineered specifically for multi-step AI agent workflows. The models are designed to improve the accuracy of agentic retrieval, providing a dedicated layer for agents querying complex knowledge bases. The flagship 8B variant immediately captured the #1 overall position on the Retrieval Embedding Benchmark (RTEB) as of July 15, 2026.

Model Variants and Architecture

The release includes three primary variants built on Transformer architectures utilizing bidirectional attention masking. All three models support a 32k context window, allowing agents to retrieve and process long-form documents, complex code bases, and extensive multi-turn histories in a single pass.

Model	Backbone / Size	Format	Target Use Case
Nemotron-3-Embed-8B-BF16	Ministral-3-8B-Instruct-2512	BF16	Precision-critical enterprise retrieval
Nemotron-3-Embed-1B-BF16	1.14 billion parameters	BF16	Standard production environments
Nemotron-3-Embed-1B-NVFP4	1.14 billion parameters	NVFP4	Ultra-high throughput inference

The models feature native multilingual support, evaluated across 34 languages including English, Arabic, Chinese, Hindi, Japanese, and Russian. NVIDIA released the collection under the OpenMDW-1.1 license, enabling commercial use.

Benchmark Performance

The models were evaluated against 16 public tasks within the RTEB. This benchmark serves as a modern industry standard designed to measure real-world retrieval accuracy for complex tasks rather than static semantic similarity. The 8B-BF16 variant secured the #1 overall rank on the leaderboard.

This release addresses a known issue called small-model collapse. Older benchmarks like MTEB struggled to distinguish the quality difference between 1B and 8B embedding models, making it difficult to evaluate the actual retrieval quality of larger architectures. The RTEB results validate the structural advantages of the 8B backbone for agent workflows.

Independent testing by Artificial Analysis recorded inference speeds of up to 454.4 tokens per second when serving Nemotron models via BLACKBOX AI, outperforming competing models by nearly 50% in throughput.

Hardware Acceleration and Ecosystem

The 1B-NVFP4 variant leverages 4-bit quantization optimized specifically for NVIDIA Blackwell (B300/GB200) architectures. This hardware acceleration delivers a significantly smaller memory footprint. When deployed locally, the NVFP4 format allows state-of-the-art retrieval capabilities to run in approximately 5GB of VRAM, making it accessible on consumer hardware such as the RTX 4060 Ti.

NVIDIA integrated the models into its broader software stack. They are available as NVIDIA NIM microservices for production inference. For custom use cases, NVIDIA open-sourced NeMo AutoModel recipes to facilitate domain adaptation, allowing teams to fine-tune the models on proprietary enterprise data. Cloud and inference deployment partners at launch include Baseten, Bitdeer AI, DeepInfra, Friendli AI, and OpenRouter.

If you build large-scale RAG systems, this release shifts the baseline for what is possible with open-weight retrieval models. Test the 1B-NVFP4 variant first to determine if its latency and memory profile meet your serving constraints before scaling up to the 8B model for queries requiring higher precision.