HomeAINVIDIA Outlines Nemotron Labs Open AI
AI

NVIDIA Outlines Nemotron Labs Open AI Model Framework

NVIDIA outlined its Nemotron Labs framework, highlighting how enterprises customize open models, datasets, and software libraries to reduce inference costs.

WHAT YOU NEED TO KNOW
  • Harvey and LangChain achieved top legal and agent accuracy on Nemotron 3 Ultra at 10 times lower cost per run than leading closed models.
  • Arcee AI ran post-trained Nemotron on the NVIDIA Blackwell platform at approximately 90 cents per million output tokens.
  • H Company's Holotron 3 Nano achieved over 76 percent accuracy on the OSWorld-Verified computer task benchmark.
  • NVIDIA made Nemotron 3 Ultra and Nemotron 3 Nano Omni accessible through build.nvidia.com alongside the Nemotron Coalition ecosystem.

NVIDIA outlined details of its Nemotron Labs initiative, showing how enterprises and organizations customize open models, datasets, and software libraries on its hardware platforms.

The company said open models like NVIDIA Nemotron allow businesses to inspect and adjust systems locally without routing proprietary information through third parties. To accelerate customization, evaluation, and governance, NVIDIA provides the NeMo suite of open libraries alongside tools from ecosystem partners including Prime Intellect and Unsloth.

Partner deployments

Several software developers have adapted Nemotron models for specific industry tasks. Legal platform Harvey post-trained Nemotron 3 Ultra on its internal benchmark, matching leading closed models at at least 10 times lower cost per run. Healthcare developer Abridge is customizing Nemotron to build a foundation model for clinical conversations, while Heidi Health uses the architecture for clinical documentation.

Developer framework LangChain adjusted prompts, tools, and middleware for Nemotron 3 Ultra without retraining the model itself. LangChain reported top agent accuracy among open models at approximately 10 times lower cost per run than leading closed alternatives.

Benchmark results and infrastructure

H Company post-trained Nemotron 3 Nano Omni on proprietary computer-use data to build Holotron 3 Nano. The resulting model reached over 76 percent accuracy on the OSWorld-Verified benchmark for computer tasks. Model builder Arcee AI post-trained Nemotron on NVIDIA's Blackwell hardware platform, achieving inference costs of roughly 90 cents per million output tokens while ranking second on PinchBench.

For regional customization, YTL AI Labs post-trained a Nemotron model specifically for the Malaysian language. Enterprise search provider Glean developed an agentic search model named Waldo, which pairs Nemotron with larger closed models to operate at lower latency with fewer required tokens.

NVIDIA established the Nemotron Coalition to aggregate shared datasets, evaluations, and domain expertise across participating organizations. Developers can access open models, weights, and datasets, or test Nemotron 3 Ultra and Nemotron 3 Nano Omni directly through build.nvidia.com.

Xentir Media
Xentir Media NewsroomSource-backed AI and technology coverage, drafted by Xentir's automated editorial system under fixed human-set rules. See our editorial policy and AI usage policy.
J
Jomon · Founder & EditorFounder and editor of Xentir Media. Sets the editorial rules the newsroom system runs under, and is accountable for its corrections. About Jomon · [email protected]
The Xentir Brief
The developments worth knowing — one useful email.
Get the Brief →
IN THE AI INDEX

Models named in this story, with their current rank on the index:

Nemotron 3 Ultra · #69 overallSee the full AI Model Rankings →