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

On July 16, 2026, NVIDIA released the Nemotron 3 Embed series, with the flagship model Nemotron-3-Embed-8B-BF16 ranking first on RTEB (Retrieval Task Evaluation Benchmark) with an NDCG@10 score of 78.5%. The model also achieved a high score of 75.5% on MMTEB Retrieval, significantly surpassing the previous Nemotron baseline.

The 1B variant Nemotron-3-Embed-1B-BF16 scored 72.4% on RTEB, reducing the error rate by 27% compared to its predecessor llama-nemotron-embed-vl-1b-v2; on MMTEB Retrieval it scored 71.0%, a 28% error reduction. Evaluations also covered ViDoRe V3 Text and LongEmbed benchmarks, using average NDCG@10 metrics, further validating the model's robustness in long-document and multilingual scenarios.

To measure the impact of retrieval on agent workflows, NVIDIA designed an experiment: using Nemotron 3 Ultra as the search agent, only replacing the embedding model, evaluating average retrieval accuracy and downstream agent token cost (estimated using GPT-5.5 pricing) on three datasets: ViDoRe V3, BRIGHT, and BrowseComp-Plus. Results show that stronger retrieval returns relevant evidence earlier, reducing the agent's repeated searches and unnecessary reasoning steps, thereby lowering token consumption.

Nemotron-3-Embed-8B-BF16 achieved both the highest average retrieval accuracy and the lowest estimated token cost across the three datasets, establishing a new frontier for agent retrieval. The 1B variant also significantly outperformed its predecessor, indicating that even on small models, improvements in retrieval quality directly translate into enhanced agent efficiency.

For high-throughput deployment, NVIDIA introduces Nemotron-3-Embed-1B-NVFP4, leveraging Blackwell architecture's native NVFP4 acceleration to quantize linear layer weights and activations to 4 bits, and using quantization-aware distillation (QAD) to recover long-sequence accuracy. On ViDoRe V3, this variant retains over 99% of the retrieval accuracy of the BF16 version, while achieving up to 2x throughput improvement and significantly reduced memory footprint.

Compared to small models like Qwen3-Embedding-0.6B and EmbeddingGemma-300M, the NVFP4 variant holds a clear advantage on the accuracy-efficiency curve. NVIDIA also provides a Rust-based NIM microservice; on GB200 and RTX PRO 6000 GPUs, for input sequence lengths of 256 and 1024, NIM performance matches or exceeds vLLM checkpoints.

Nemotron-3-Embed-8B-BF16 is based on the Ministral-3-8B-Instruct-2512 backbone, converting its causal decoder into a bidirectional encoder for full-sequence retrieval. Training consists of two stages: first, contrastive pretraining on a mixture of web sources and synthetic text pairs; then, fine-tuning on multi-domain, multilingual retrieval datasets covering law, finance, healthcare, business, and education.

The 1B model transfers knowledge from the 8B model through distillation and quantization techniques. NVIDIA open-sources the weights, datasets, and NeMo AutoModel fine-tuning/distillation recipes, allowing teams to perform domain adaptation and model compression on their own data.

The Nemotron 3 Embed series received broad ecosystem support on its release day: model weights are available on Hugging Face, deployable via NVIDIA NIM microservices, and compatible with the vLLM inference framework. The 32K context window supports long documents, codebases, and multi-turn agent history retrieval, reducing truncation loss.

Multilingual and code retrieval capabilities cover global enterprise data and technical documentation. NVIDIA emphasizes that open weights and recipes give teams full control over inspection, tuning, and deployment of retrieval models, suitable for production-grade RAG, agent retrieval, code retrieval, and agent memory scenarios.