NVIDIA has made a significant contribution to the open-source AI landscape with the release of Nemotron 3 Embed, an innovative embedding collection designed to advance retrieval-augmented generation (RAG) and other downstream tasks. The release, announced on July 15 and 16, 2026, includes three open checkpoints: Nemotron-3-Embed-8B-BF16, Nemotron-3-Embed-1B-BF16, and Nemotron-3-Embed-1B-NVFP4. Each of these models is built to balance performance, efficiency, and accessibility.
Performance and Efficiency at the Forefront
The standout model in the collection is the Nemotron-3-Embed-8B-BF16, which has achieved a top rank on the Retrieval Evaluation Benchmark (RTEB) with an average NDCG@10 score of 78.46. This performance places it at the #1 position, underscoring its effectiveness in information retrieval tasks. The 1B models, on the other hand, were developed using advanced techniques like ModelOpt NAS pruning and COS+MSE distillation from the 8B teacher model, ensuring they maintain high accuracy while being more resource-efficient.
Scalability and Speed Optimizations
One of the most notable features of the Nemotron 3 Embed collection is its ability to process long sequences efficiently. All models support 32,768-token inputs under the OpenMDW-1.1 framework, which is crucial for handling complex, real-world data. Additionally, the Nemotron-3-Embed-1B-NVFP4 variant maintains over 99% of the BF16 retrieval accuracy while delivering up to twice the throughput on NVIDIA Blackwell hardware. This makes it an ideal choice for applications where speed and efficiency are paramount without compromising on performance.
The release of Nemotron 3 Embed reflects NVIDIA’s ongoing commitment to democratizing AI through open-source tools. By providing high-performing, scalable, and efficient models, NVIDIA is enabling developers and researchers to build more powerful and accessible AI systems.
