NVIDIA's New AI Search Models Top Global Rankings and Cut Agent Costs

Key points

• NVIDIA released three Nemotron 3 Embed models on 15 July 2026, all free to use commercially.
• The flagship 8B model ranked first on RTEB, the main global leaderboard for measuring how well AI finds relevant information.
• The smaller 1B model cut retrieval error rates by 27 to 28 percent compared to NVIDIA's previous 1B model.
• The hardware-optimised 1B variant delivers up to twice the processing speed on NVIDIA's Blackwell chips while keeping accuracy above 99 percent of the standard version.
• All three models are available immediately on Hugging Face.

When you ask an AI assistant a question, it usually does not guess the answer from memory. It searches a library of documents, pulls out the most relevant ones, and then writes a reply based on what it found. That search step is called retrieval, and how well it works shapes everything that follows.

NVIDIA Corporation, the chip maker that also builds AI software tools, released a new family of retrieval models this week called Nemotron 3 Embed. An embedding model, to explain the term plainly, is an AI that converts words and sentences into numbers so that a computer can quickly compare how closely related two pieces of text are. Good embedding models find the right documents fast. Poor ones return noise, waste the AI agent's time, and inflate computing costs.

The flagship model, Nemotron-3-Embed-8B-BF16, took the top spot on RTEB (the Retrieval Text Embedding Benchmark, the field's main public accuracy contest) as of 15 July 2026, scoring 78.5 percent. That matters because RTEB tests retrieval across many languages, document types, and tasks, so a top score there is hard to game with narrow training.

Does better retrieval actually save money?

Yes, and the savings can be substantial. NVIDIA ran a telling experiment: they connected their new embedding models to an AI agent, a software program that can carry out multi-step tasks on its own, and measured how many tokens, the units of text an AI model reads and writes, the agent needed to complete a search task. More accurate retrieval returned useful results earlier, so the agent needed fewer repeated searches and fewer reasoning steps. Fewer steps means lower token counts, and lower token counts mean lower bills.

The 8B model produced both the highest accuracy and the lowest estimated token cost across three standard benchmarks. That is not a trade-off; it is simply a better model doing less unnecessary work.

For organisations that cannot afford to run an 8-billion-parameter model at scale, NVIDIA also released a 1-billion-parameter version, Nemotron-3-Embed-1B-BF16. It scores 72.4 percent on RTEB, still strong, and cuts error rates by 27 percent against its predecessor. A third variant, Nemotron-3-Embed-1B-NVFP4, uses a compressed number format called NVFP4 on NVIDIA's Blackwell hardware to run at up to twice the speed of the standard version while losing less than one percent of accuracy.

All three models support documents up to 32,000 tokens long, handle multiple languages, and work with code repositories. NVIDIA has published the model weights, training data details, and fine-tuning recipes, so teams can adapt the models to their own documents without starting from scratch.

These are company benchmarks and NVIDIA's own evaluations. Independent clinical-grade trials do not apply here, but the RTEB leaderboard is maintained independently, and the top ranking is a meaningful external signal. Developers wanting to test the models on their own data can download them today, as first noted by Hugging Face, where all three are listed.