主なポイント
- NVIDIA has announced NVIDIA Nemotron 3 Embed, a new suite of open embedding models for enterprise search and AI agent retrieval.
- Boomi is supporting Nemotron 3 Embed 1B within Boomi Knowledge Hub.
- By adopting open models, enterprises can reduce token waste and lower AI operational costs, maximizing the performance of their private AI ecosystems.
Boomi’s position on enterprise AI is simple. Our CEO, Steve Lucas, laid it out in his recent whitepaper “Economics Always Wins”: every serious enterprise will end up running private AI. But in practice, winning on economics isn’t about shutting out the massive frontier models that excel at complex reasoning — it’s about optimizing how we feed them. The smartest enterprises are realizing that using costly external models to sift through raw, unstructured local data is highly inefficient. Instead, they are building private AI foundations that handle the heavy lifting of data indexing and search locally. That is why open retrieval is a game-changer: by using highly accurate, open embedding models to pre-filter and pinpoint exact context, organizations can feed highly refined data downstream, drastically reducing wasted tokens and maximizing the ROI of their entire AI ecosystem.
Today, that argument reaches the layer where enterprises feel it most: retrieval. NVIDIA announced NVIDIA Nemotron 3 Embed, open embedding models in various sizes for enterprise search and agentic retrieval. Boomi is supporting Nemotron 3 Embed 1B in Boomi Knowledge Hub, our AI-ready data service.
Retrieval Decides Agent Quality and Cost
AI agents don’t retrieve information once. They retrieve continuously, to plan, use tools, check memory, and stay grounded in facts. Every retrieval that returns the wrong context wastes tokens, raises costs, and weakens the answer. At agent scale, retrieval quality is a cost line, not just an accuracy metric. The economics compound with every query, and embedding models are where those economics start.
Boomi Knowledge Hub ingests enterprise data from the systems where it lives, indexes it for retrieval, and serves relevant, cited context to AI agents at query time, so agents answer from your organization’s knowledge instead of guessing. Every chunk it indexes and every query an agent sends passes through an embedding model. A more accurate model means agents find the right context in fewer attempts, with fewer wasted tokens downstream.
What Nemotron 3 Embed Brings
Nemotron 3 Embed gives enterprises a practical way to balance retrieval accuracy, throughput, and cost. Nemotron 3 Embed 8B tops the RTEB leaderboard across both open and closed embedding models, making it the option for maximum retrieval quality. Nemotron 3 Embed 1B is built for high-volume workloads that need strong accuracy with low latency and cost. It is available in BF16 and NVFP4 variants; with NVFP4 on NVIDIA Blackwell, the 1B model doubles throughput of BF16 while retaining 99% of BF16 retrieval accuracy.1 That makes it fast and inexpensive enough for AI agents that retrieve constantly. Examples include customer service agents, scheduling agents, dashboard agents, or any other agents that rely on information in a knowledge base, on the internet, or other sources that are frequently updated.
Just as important: the models are open. Open weights, open datasets, and open recipes mean enterprises can inspect, fine-tune, and deploy them on their own terms. No black box, no lock-in.
We didn’t take that on faith. With early access, Boomi’s engineering team evaluated Nemotron 3 Embed 1B on accuracy and throughput against the embedding models we’ve historically used across the platform, testing with our internal Boomi Answers query sets, which reflect the diverse questions real users ask, and with standard public retrieval benchmarks. That’s the bar for any model we put in front of customers: it has to earn its place on our workloads, not just on a leaderboard. This rigor ensures our customers experience high quality retrieval in every agent interaction, in both the answers they get and the tokens they pay for. Nemotron 3 Embed 1B delivered competitive retrieval accuracy — the profile that matters most for agents that retrieve constantly — at markedly better throughput and cost than the closed options we tested.
Openness and Choice, Managed for You
Boomi Knowledge Hub is designed to make model choice practical. It maintains an extensible catalog of embedding models, so customers can select the model that fits their accuracy, cost, and language requirements. Nemotron 3 Embed extends that catalog with the leading open option, and Boomi manages the ingestion, indexing, and retrieval pipeline around it. You get the economics of open models with the simplicity of a managed service: no embedding infrastructure to build and a streamlined process for replacement when a better model arrives.
Retrieval is becoming the default step in every agent decision. Open, accurate, efficient embedding models served through a governed context layer are how enterprises keep both quality and cost under control. Economics always wins; this is what it looks like winning.
Check out NVIDIA’s blog on NVIDIA Nemotron 3 Embed to learn more, and explore Boomi Knowledge Hub through Boomi’s Early Access program on the Boomi Community.
–
1. http://huggingface.co/blog/nvidia/nemotron-3-embed-wins-rteb