Nvidia releases open Nemotron models for private workplace AI

Nvidia has introduced open-weight Nemotron AI models designed to let businesses run agents on their own infrastructure and workers use them locally on PCs, according to TechCrunch. The releases extend the lineup from enterprise automation to desktop assistants that can operate without sending workloads to the cloud.

The enterprise model, Nemotron 3.5 Super VL, comes with open weights, datasets and training recipes. Businesses can further train the 120-billion-parameter model on their own data and deploy it where their security and data-residency requirements dictate, Nvidia said in its launch blog. Its document capabilities include reading text, tables and charts in webpages and PDFs.

Nvidia named specific business tasks already being used to customize the model. Scale AI has post-trained it for EY’s merger-and-acquisition due diligence agent and Skydance’s vendor credit finance agent, the company said. Ramp is customizing it for financial-document intelligence, while R1 is doing so for medical-claims analysis and revenue-cycle management.

On the desktop, Nemotron 3.5 Bolt 74B-A9B 24GB is designed to fit in 24GB of memory and run alongside other applications on supported Nvidia hardware. The company says users can put it to work researching the web, writing and debugging code locally. The smaller Bolt E4B 4GB targets conversational and physical agents, including local voice assistants.

Getting businesses comfortable with software running locally again, rather than only through cloud or web services, remains an adoption obstacle, Omdia analyst Alexander Harrowell told TechTarget.

Super VL and Bolt 74B-A9B are available now. Nvidia’s launch blog schedules Bolt E4B for Oct. 23.