NVIDIA Nemotron 3 Embed Leads RTEB Retrieval Rankings

NVIDIA has released Nemotron 3 Embed, a family of open, commercially available embedding models built to improve the retrieval stage of AI systems. Embedding models convert text, code, and other supported content into numerical representations so that a system can find material with similar meaning. This capability is central to retrieval-augmented generation (RAG), search agents, code assistants, and agent memory, where the quality of the retrieved evidence can directly affect the quality, speed, and cost of the final response. The flagship Nemotron-3-Embed-8B-BF16 model ranks first overall on the RTEB multilingual retrieval leaderboard as of July 15, 2026.

The collection includes an 8-billion-parameter model intended to establish the highest retrieval-quality tier, as well as smaller 1-billion-parameter BF16 and NVFP4 variants designed for less expensive, higher-throughput production deployments. NVIDIA presents the family as covering different points on the accuracy-versus-efficiency curve rather than forcing every organization to deploy the largest model. The 8B BF16 version is aimed at use cases where retrieval quality is the main priority, while the 1B models bring the same retrieval-centered design into environments with tighter latency, memory, and infrastructure constraints.

Nemotron 3 Embed is being released with open weights, datasets, and training recipes. This gives development teams the ability to inspect the models, deploy them on their own infrastructure, tune them for specialized domains, and fine-tune them using private or industry-specific data. NVIDIA also provides NeMo AutoModel recipes for domain adaptation, knowledge distillation, and model compression. These resources are meant to help organizations adapt a general retrieval model to their own documents, terminology, codebases, or workflows without depending entirely on a closed hosted service.

The models support a context window of up to 32,000 tokens. That capacity allows the retriever to process longer documents, larger sections of source code, and extended multi-turn agent histories while discarding less information through truncation. The family also supports multilingual and code retrieval, addressing organizations whose searchable data spans multiple languages, technical documentation, and repositories containing many related files. Immediate ecosystem support includes distribution through Hugging Face, deployment as an NVIDIA NIM microservice, compatibility with vLLM, and access through AI cloud and inference partners.

NVIDIA evaluated the models across retrieval quality, downstream efficiency in agent workflows, and practical deployment tradeoffs. The reported tests cover RTEB, ViDoRe V3 Text, MMTEB Retrieval, and LongEmbed, using average NDCG@10 as a retrieval metric. NDCG@10 measures whether relevant results appear near the top of the first ten returned items, giving greater credit when the best evidence is ranked earlier. Nemotron-3-Embed-8B-BF16 scored 78.5% on RTEB and 75.5% on MMTEB Retrieval, with its RTEB result placing it at the top of the cited leaderboard.

The smaller Nemotron-3-Embed-1B-BF16 model scored 72.4% on RTEB. NVIDIA says this represents a 27% reduction in error rate compared with its previous 1B model, llama-nemotron-embed-vl-1b-v2. On MMTEB Retrieval, the new 1B BF16 model scored 71.0%, corresponding to a reported 28% error-rate reduction over that predecessor. These results suggest that the compact model retains a substantial portion of the flagship model’s retrieval performance while requiring a smaller deployment footprint.

The company also tested how the choice of embedding model affects an AI agent after retrieval. In this setup, a search agent powered by Nemotron 3 Ultra used different embedding models to locate evidence across ViDoRe V3, BRIGHT, and BrowseComp-Plus. The evaluation compared average retrieval accuracy with the estimated number and cost of downstream agent tokens per query. NVIDIA estimated that cost from Nemotron 3 Ultra input and output token counts using a GPT-5.5 pricing formula; it is therefore a modeled cost comparison rather than a direct billing measurement for the embedding models themselves.

The underlying argument is that better retrieval can reduce work elsewhere in an agentic pipeline. If relevant evidence appears earlier, an agent may need fewer repeated searches, fewer reasoning turns, and less inspection of irrelevant context. In NVIDIA’s evaluations, the 8B model combined the highest average retrieval accuracy with the lowest estimated downstream agentic token cost among the compared Nemotron 3 Embed variants. The result highlights that a larger retriever can potentially lower total system cost when its improved ranking prevents an expensive reasoning model from doing unnecessary work.

For deployments where serving throughput and memory usage are more important, NVIDIA offers Nemotron-3-Embed-1B-NVFP4. It uses the NVFP4 4-bit numerical format for the weights and activations of linear layers and is optimized for NVIDIA Blackwell hardware. Lower-precision computation reduces the resources required for inference, but it can also reduce model quality. To limit that loss, NVIDIA applies Quantization-Aware Distillation, a training method that teaches the compressed model to preserve the behavior of a more accurate version, including on long input sequences.

NVIDIA reports that native NVFP4 inference on Blackwell can deliver up to twice the throughput of BF16 in high-throughput, low-latency retrieval serving. The company also says the NVFP4 model retains more than 99% of the BF16 variant’s retrieval accuracy. Its efficiency comparison on ViDoRe V3 includes smaller open embedding baselines such as Qwen3-Embedding-0.6B and EmbeddingGemma-300M. Together, the releases position retrieval quality as a system-level optimization: a stronger embedding model may improve answer grounding while reducing the searches, reasoning steps, token use, and infrastructure resources required by the wider agent.