Nemotron Labs: How Open Models Give Enterprises and Countries AI That Is Trustworthy, Controllable, and Customizable

NVIDIA has launched the Nemotron Labs article series, dedicated to how open models help businesses build specialized AI systems. It notes that competitive advantage in AI is increasingly achieved through how organizations use available models, rather than which specific model they choose. Open models, unlike closed ones, provide full control and customization capabilities: enterprises can inspect, fine-tune, and improve AI for their tasks. Effective agentic applications are built as model systems, where open models work alongside leading frontier models. Examples of companies already customizing Nemotron include: Abridge adapting the model for clinical dialogues; Glean creating the agentic search model Waldo; H Company building Holotron 3 Nano; Harvey fine-tuning Nemotron for legal tasks with accuracy on par with closed alternatives at a 10x lower cost; Heidi Health achieving frontier-level quality without frontier-scale costs; and YTL AI Labs training a model for the Malay language. The NVIDIA NeMo library suite accelerates customization and evaluation. Partners Prime Intellect and Unsloth assist with post-training. LangChain set up an agentic wrapper for Nemotron 3 Ultra and achieved the best accuracy among open models. Arcee AI, using the NVIDIA Blackwell platform, achieved costs of approximately 90 cents per million output tokens — about 20x cheaper than closed alternatives. The NVIDIA Nemotron coalition brings together developers for collaborative improvements to the open model.