NVIDIA released three open-weight embedding models on July 15, detailed in a post on the Hugging Face blog under the name Nemotron 3 Embed. The flagship, an 8 billion parameter model, ranked first overall on RTEB, a leaderboard that scores retrieval-focused embedding models. Retrieval is the layer that decides what text an AI agent actually reads before it reasons, and NVIDIA is positioning this release as infrastructure for that layer rather than another chatbot upgrade.

Nemotron-3-Embed-8B-BF16 scored 78.5 percent on RTEB and 75.5 percent on MMTEB Retrieval, according to NVIDIA’s own published evaluation. The two smaller models, a 1B version in standard precision and a 1B version quantized to NVIDIA’s 4-bit NVFP4 format for Blackwell GPUs, are built for production serving where cost and latency outweigh the last point of accuracy. The BF16 1B model scored 72.4 percent on RTEB and 71.0 percent on MMTEB Retrieval, cutting the error rate by 27 and 28 percent respectively against NVIDIA’s prior 1B embedding model. The NVFP4 variant retains more than 99 percent of that accuracy while roughly doubling throughput on Blackwell hardware, per NVIDIA’s benchmarks.

Those figures all come from NVIDIA’s own test suite. The release does not include independent third-party benchmark results, so the RTEB rank should be read as a strong internal claim rather than a neutral audit.

The economics case is the more interesting argument. NVIDIA built a search agent on its own Nemotron 3 Ultra model and measured retrieval accuracy against downstream token cost across three benchmarks: ViDoRe V3, BRIGHT, and BrowseComp-Plus. A weak retriever forces an agent to re-query, inspect more documents, and drag noise forward into whatever reasoning follows, all of which shows up as token spend. The 8B model produced both the highest retrieval accuracy and the lowest downstream token cost of any embedding model NVIDIA tested in that setup.

Open weights are the strategic bet underneath the benchmark numbers. Nemotron 3 Embed ships with open weights, training data, and fine-tuning recipes through NVIDIA’s NeMo AutoModel toolkit, a structurally different offer than the closed embedding APIs sold by OpenAI and Cohere, which currently anchor much of enterprise RAG. A team that can fine-tune the retriever on its own corpus is not stuck with a vendor’s general-purpose ranking. NVIDIA published one such result: fine-tuning the 1B model on its own documentation lifted NDCG@10 from 56.7 percent to 63.3 percent and Recall@5 from 56.1 percent to 62.8 percent.

Several companies building retrieval and agent-memory products said they are already testing the models, including Automation Anywhere, Boomi, IBM, Palantir, ServiceNow, and the memory startup Mem0. Mem0 published its own number: the 1B model scored 80.38 percent against Qwen3-Embedding-0.6B’s 78.71 percent on LongMemEval, a memory-retrieval benchmark, in a test run outside NVIDIA’s chosen comparisons. That is an early but useful signal, since Mem0 had no incentive to make NVIDIA look good.

For any team running RAG or agent memory on a closed embedding API, Nemotron 3 Embed is now a fine-tunable, self-hostable option worth benchmarking before the next renewal, especially if repeated agent queries are already a visible line item in the token bill.

NVIDIA published these results in a post on the Hugging Face blog on July 15, 2026.