NVIDIA has introduced a series of open models, including the Nemotron family, designed to give enterprises greater control over AI systems. These models are intended to be tailored to business requirements, allowing organizations to build AI that is controllable, trustworthy, and aligned with their unique workflows. By providing full access to the model, NVIDIA aims to empower businesses to improve accuracy, test against their own data, and refine AI systems without relying on third-party infrastructure.
Customization of open models like Nemotron allows enterprises to adapt AI to their specific domains. For example, Abridge is developing a foundation model for clinical conversations, while Glean has created an agentic search model that pairs Nemotron with larger closed models to achieve lower latency and fewer tokens. Additionally, H Company achieved over 76% accuracy on a computer task benchmark by post-training Nemotron 3 Nano Omni with proprietary data, and Harvey matched leading closed models in legal tasks at a cost 10 times lower. These examples highlight how open models can be fine-tuned for industry-specific applications, improving performance and reducing costs.
The source emphasizes that open models remove barriers that closed models impose, allowing enterprises to inspect, tune, and improve AI systems. This approach supports the development of agentic AI applications, where multiple models work together to handle complex tasks efficiently. The NVIDIA NeMo suite of open libraries further accelerates customization and evaluation, enabling partners to build scalable AI solutions. By post-training Nemotron on the Blackwell platform, Arcee AI achieved inference costs of about 90 cents per million output tokens, significantly lower than comparable closed models.
Source: nvidia
