NVIDIA Nemotron 3 Embed Claims Top RTEB Ranking with Open Retrieval Models

NVIDIA has released Nemotron 3 Embed, a collection of open and commercially available embedding models aimed at improving retrieval quality across production-scale retrieval-augmented generation (RAG), agentic retrieval, code retrieval, and agent memory applications. The collection features three open models, led by an 8B flagship that claims the top position on the RTEB leaderboard, alongside efficient 1B variants designed for large-scale deployment.

Top-Tier Retrieval Performance

The flagship Nemotron-3-Embed-8B-BF16 model ranks first on RTEB as of July 15, 2026, scoring 78.5% on RTEB and 75.5% on MMTEB Retrieval. The smaller Nemotron-3-Embed-1B-BF16 brings much of that retrieval quality into a more compact footprint, scoring 72.4% on RTEB — a 27% error rate reduction over its 1B predecessor — and 71.0% on MMTEB Retrieval, representing a 28% error rate reduction.

Beyond RTEB, NVIDIA evaluated the models across ViDoRe V3 Text, MMTEB Retrieval, and LongEmbed using average NDCG@10. The 8B model establishes the collection's quality ceiling, while the 1B variants extend the same retrieval-focused design to lower-cost, higher-throughput settings.

Agentic Efficiency and Deployment Flexibility

To assess retrieval in agentic settings, NVIDIA paired the embedding models with a search agent powered by Nemotron 3 Ultra, measuring both retrieval accuracy and estimated downstream agentic token cost per query across ViDoRe V3, BRIGHT, and BrowseComp-Plus. The results show that stronger retrieval reduces downstream token costs, as more accurate retrievers return relevant evidence earlier, helping agents complete tasks with fewer repeated searches and reasoning turns. The 8B model delivered both the highest average retrieval accuracy and the lowest estimated downstream token cost across these benchmarks.

For high-throughput deployments, the Nemotron-3-Embed-1B-NVFP4 variant uses native NVFP4 acceleration on NVIDIA Blackwell architectures. This approach quantizes weights and activations of linear layers to 4-bit format and employs Quantization-Aware Distillation to recover accuracy for long input sequences. The NVFP4 variant retains over 99% of BF16 retrieval accuracy while reducing memory footprint, and delivers up to 2x higher throughput than BF16 on Blackwell hardware.

NVIDIA also released an optimized NIM microservice for the 1B model. The Rust-based Nemotron 3 Embed NIM matches or outperforms the vLLM checkpoint on NVIDIA GB200 and RTX PRO 6000 GPUs across input sequence lengths of 256 and 1024 tokens.

Architecture and Training Pipeline

The 8B model adapts the Ministral-3-8B-Instruct-2512 backbone by converting its causal decoder into a bidirectional encoder for full-sequence retrieval. Training involved contrastive pre-training on web-sourced and synthetic text pairs, followed by fine-tuning on curated multilingual retrieval datasets spanning legal, finance, medical, business, and education domains. Earlier 8B teacher checkpoints from the same development line were used to distill the efficient 1B variants.

The 1B model was not trained from scratch. NVIDIA first applied the bidirectional adaptation recipe to the Ministral-3-3B-Instruct-2512 backbone to create a 3B retriever base, then compressed it through two rounds of structured pruning and distillation. The first round used NVIDIA ModelOpt's mcore_minitron Neural Architecture Search engine to compress the 3B model to a 2B intermediate, searching across hidden width, FFN size, attention heads, and depth under a strict parameter budget. The 2B model was then distilled from an 8B teacher using combined cosine distance and mean squared error losses on multilingual in-domain retrieval data.

A second round of pruning and distillation compressed the 2B intermediate to the final 1.14B model. Final training used a progressive two-stage context-scaling schedule: the first stage focused on broad multilingual alignment at 1024-token context length, while the second expanded to 4096 tokens with long-context synthetic and reasoning datasets.

Enterprise Adoption and Availability

Nemotron 3 Embed ships with a production-ready feature set including a 32k context window for long documents and multi-turn agent histories, multilingual and code retrieval support, open weights, datasets, and fine-tuning recipes. The models are available immediately on Hugging Face with support for SentenceTransformers, Transformers, and vLLM, as deployable NVIDIA NIM microservices, and through AI Cloud and inference partners including Baseten, Bitdeer AI, DeepInfra, Friendli AI, and OpenRouter.

NVIDIA has also open-sourced NeMo AutoModel training recipes for fine-tuning and distillation. In one example, fine-tuning Nemotron-3-Embed-1B-BF16 on the NV Docs evaluation improved NDCG@10 from 56.7% to 63.3% and Recall@5 from 56.1% to 62.8%.

Numerous enterprises are already evaluating the models. Automation Anywhere reported promising early results for question answering. Boomi highlighted the flexibility of having both 1B and 8B variants. IBM tested the model in a proof-of-concept on watsonx.data. Mem0 is evaluating it for AI agent memory. Palantir is collaborating on edge retrieval workloads. ServiceNow is testing retrieval through its documentation. turbopuffer integrated the models into its native embeddings service. You.com reported a significant performance leap in its re-ranking stack. Zep found the 1B model ranked first across all memory retrieval tasks in internal benchmarks. Zoom is evaluating the models for its enterprise agentic search contextual layer.

As enterprises race to build more capable agentic AI systems, retrieval quality is becoming a critical bottleneck — and NVIDIA's Nemotron 3 Embed addresses this challenge with open, flexible models that balance accuracy and efficiency. Have you tried these new embedding models in your projects? Share this article with your network and join the conversation about the future of retrieval-augmented AI.

Source: Hugging Face Blog