NVIDIA Releases Nemotron 3 Embed for Production RAG

NVIDIA released its Nemotron 3 Embed model collection on July 17, 2026, targeting production-scale retrieval-augmented generation (RAG), agentic retrieval, code retrieval, and agent memory. The collection features three open checkpoints: Nemotron-3-Embed-8B-BF16, Nemotron-3-Embed-1B-BF16, and Nemotron-3-Embed-1B-NVFP4. These models are transformer encoders trained with bidirectional attention masking, with the final embedding derived from average pooling over token-level representations. All checkpoints support a maximum sequence length of 32,768 tokens and have been evaluated across 34 languages. The models are released under the OpenMDW License Agreement, version 1.1 (OpenMDW-1.1).

The Nemotron-3-Embed-8B-BF16 checkpoint, with approximately 8 billion parameters and a 4096 embedding dimension, achieved the top rank on the Retrieval Embedding Benchmark (RTEB) as of July 17, 2026. It scored 78.46 on RTEB, 60.60 on ViDoRe-V3 text, and 75.45 on MMTEB (Retrieval). The Nemotron-3-Embed-1B-BF16 model, with 1.14 billion parameters and a 2048 embedding dimension, scored 72.38 on RTEB, 57.74 on ViDoRe-V3 text, and 71.04 on MMTEB. The Blackwell-optimized Nemotron-3-Embed-1B-NVFP4 variant achieved a score of 72.00 on RTEB, demonstrating 99.5% retention compared to its BF16 counterpart.

Notably, the 1B models were built through a compression pipeline, not from smaller training runs. The parent model, nemotron-3-embed-3b, was pruned to 2B and then distilled across two iterative rounds using NVIDIA ModelOpt. The 8B model is built upon Mistral-3-8B-Instruct-2512, while both 1B variants utilize Mistral-3-3B-Instruct-2512 as their base. The 1B models show a significant gain of 10.4 RTEB points over the prior-generation baseline, llama-nemotron-embed-vl-1b-v2.