Nvidia touts Nemotron for specialist business AI uses

Nvidia has outlined how organisations are customising its open Nemotron AI models for specialised business uses, including healthcare, legal work, search and language localisation.

The update is part of a broader push to position open models as tools companies can inspect, tune and run for their own needs, rather than rely solely on closed systems.

Nvidia argued that businesses are placing greater emphasis on adapting models to their workflows, internal knowledge and accuracy standards. It said this shift is changing competition in artificial intelligence from choosing a model to shaping one for a specific task.

According to Nvidia, specialised AI systems such as autonomous agents and task-focused applications depend on access to the underlying model so teams can tune them with proprietary data and test them against business outcomes. The company contrasted this with closed models, which it said can limit what customers are able to inspect or improve.

Sector use

Several companies cited by Nvidia are using Nemotron in narrowly defined settings. In healthcare, Abridge is customising Nemotron to build a foundation model for clinical conversations, while Heidi Health is using the model family for clinical documentation.

In legal services, Harvey post-trained Nemotron 3 Ultra on its own legal benchmark. Nvidia said the work reached accuracy levels comparable with leading closed models on complex legal tasks while cutting the cost per run by at least a factor of 10.

Search company Glean built an agentic search model called Waldo that pairs Nemotron with larger closed models. Nvidia said this approach reduced latency and token use in enterprise search.

Elsewhere, H Company built Holotron 3 Nano by post-training Nemotron 3 Nano Omni on proprietary computer-use data. Nvidia said the model achieved more than 76% accuracy on the OSWorld-Verified benchmark for computer tasks and matched other frontier models at lower cost.

Local language work also featured in the examples. YTL AI Labs post-trained a Nemotron model for the Malaysian language, aiming to make locally customised AI available to Malaysia's developer community, according to Nvidia.

Control and cost

Nvidia said open models appeal to sectors with strict requirements for data handling and accuracy, including healthcare and legal services. It argued that these organisations need visibility into training, performance and how improvements are made when models fall short.

Open models also allow teams to run private evaluations against their own standards and create reinforcement learning environments matched to their workflows, Nvidia said. That can be done without routing proprietary data through a third party.

Much of Nvidia's argument focused on economics as well as control. It said that when models are tuned for a specific domain or harness, they can run more efficiently and lower inference costs.

The company pointed to LangChain's work on its Deep Agents harness for Nemotron 3 Ultra. Nvidia said LangChain adjusted prompts, tools and middleware without retraining the model, achieving top agent accuracy among open models at about 10 times lower cost per run than leading closed alternatives.

Nvidia also cited Arcee AI, which post-trained Nemotron on Nvidia's Blackwell platform. It said the work brought inference costs to about 90 US cents per million output tokens, roughly 20 times cheaper than comparable closed frontier models, while placing second on the PinchBench ranking.

Wider ecosystem

Alongside the model examples, Nvidia highlighted a broader effort to build an ecosystem around Nemotron. It said the Nvidia Nemotron Coalition is intended to bring model builders and developers together to improve the models through shared data, evaluations and domain expertise.

Nvidia also pointed to support from partners including Prime Intellect and Unsloth for enterprises building post-training pipelines on Nemotron. It said the Nvidia NeMo suite of open libraries is being used for model customisation, evaluation, agent optimisation and governance.

Nvidia's message is that open-weight models are becoming part of mixed AI systems rather than standing alone. In this approach, reasoning-heavy models can handle planning while smaller tuned models manage specialist tasks, giving organisations a way to balance accuracy, flexibility and running costs.

The examples suggest businesses are testing that model across sectors where domain knowledge, data control and operating costs matter as much as raw benchmark scores.