NVIDIA Releases Nemotron 3 Embed, an Open Embedding Collection With an 8B Model Topping the RTEB Leaderboard

NVIDIA has released Nemotron 3 Embed, an open collection of embedding models designed to power production-scale retrieval-augmented generation (RAG), agentic retrieval, code retrieval, and agent memory. The release, detailed by MarkTechPost on July 17, 2026, arrives as the layer that decides which passages an AI agent ever sees becomes a competitive battleground, a trend this publication's breaking AI news desk has followed as enterprises move from chatbots to systems that act on retrieved knowledge.

Embedding models decide which passages an agent ever sees, which makes them a quiet but decisive chokepoint in the generative AI stack. NVIDIA built Nemotron 3 Embed to strengthen that layer. The collection includes three open checkpoints: Nemotron-3-Embed-8B-BF16, an accuracy-first option; Nemotron-3-Embed-1B-BF16, which carries the same design into a smaller footprint; and Nemotron-3-Embed-1B-NVFP4, a Blackwell-optimized 4-bit variant. All three are transformer encoders trained with bidirectional attention masking, with the final embedding produced by average pooling over token-level representations. Each supports a maximum sequence length of 32,768 tokens.

Topping the Leaderboard

The headline result is that Nemotron-3-Embed-8B-BF16 ranks number one overall on RTEB, the Retrieval Embedding Benchmark, as of July 17, 2026, across its 16 public tasks. Scores are measured as average NDCG@10 at a model sequence length of 4,096. NVIDIA evaluated each model across 34 languages, and all three checkpoints are released under the OpenMDW License Agreement version 1.1, making them freely available for developers to download, run, and modify.

Notably, the model bases are drawn from Mistral. The 8B checkpoint is built on Ministral-3-8B-Instruct-2512, while both 1B variants use Ministral-3-3B-Instruct-2512, according to the MarkTechPost report. NVIDIA's decision to build on Mistral's foundation rather than a purely in-house architecture reflects how fluid model development has become, with open bases routinely repurposed and retrained for specialized tasks.

Compression, Not a Smaller Training Run

Two performance gaps stand out. The 1B model gains 10.4 RTEB points over the prior-generation llama-nemotron-embed-vl-1b-v2 baseline. Separately, the NVFP4 4-bit checkpoint costs only 0.38 RTEB points against its BF16 parent, retaining roughly 99.5% of its retrieval accuracy. That near-lossless compression is particularly significant for teams that need to run embeddings at scale, where memory and inference costs often dominate.

Those 1B scores were achieved through a compression pipeline rather than a smaller training run. The parent, nemotron-3-embed-3b, was pruned and distilled across two iterative rounds. First, the 3B parent was pruned to 2B using NVIDIA ModelOpt's mcore_minitron Neural Architecture Search, which searches over hidden width, FFN size, attention heads, and depth, then picks the best candidate from the top-10 Pareto front. A 50,000-sample in-domain calibration corpus scored those candidates.

Next, the 2B model was distilled from the fine-tuned 8B embedding teacher, combining cosine distance loss and mean squared error loss over a multilingual, in-domain data blend. The same procedure was repeated to produce the 1.14B checkpoint, demonstrating that the smaller variants inherit much of the larger model's capability through structured compression rather than independent training.

Built for Blackwell Efficiency

Compression continues into the serving format. Quantization targeted weights and activations of linear layers only, using the NVFP4 data type, with the research team relying on nvidia-modelopt v0.45.0. Quantization-Aware Distillation followed, primarily to recover accuracy on long inputs. Calibration used 512 samples, split between 256 queries and 256 passages from the cnn_dailymail dataset, while QAD training drew on 20,000 samples.

The practical payoff is efficiency on NVIDIA's latest hardware. The research team reports that NVFP4 on Blackwell delivers up to 2x higher throughput than BF16 while retaining more than 99% of BF16 retrieval accuracy. The NVFP4 model card also documents dynamic embedding sizes, allowing developers to slice the 2,048-dimensional vector down to 1,024 or 512 dimensions and re-normalize afterward, trading a little accuracy for lower storage cost. That flexibility matters for vector databases, where storing billions of high-dimensional vectors is expensive.

Why Embeddings Matter Now

Embedding models have become a quiet chokepoint in the generative AI stack. Every retrieval-based agent, enterprise search tool, and code assistant depends on them to fetch the right context before a large language model generates a response. A stronger, cheaper embedding model can directly improve answer quality and cut operating costs, which is why vendors from OpenAI to Cohere to open-source collectives have rushed to release new families.

By open-sourcing a top-ranked collection under a permissive license and pairing it with Blackwell-tuned quantization, NVIDIA is reinforcing its strategy of seeding the broader AI ecosystem with tools that run best on its own silicon. For developers building RAG pipelines or agent memory systems, Nemotron 3 Embed offers a fresh, freely available option that pushes the state of the art on a widely watched benchmark, while the NVFP4 variant gives teams a credible path to cutting inference cost without sacrificing the retrieval quality their applications depend on.

Stay Ahead of AI

Open embedding models like Nemotron 3 Embed are quietly reshaping how AI agents find and use information. For more on the models, releases, and research moving the field forward, follow our latest AI developments as they break.

Read more AI news ->