Nvidia launches new open models to run agents in the enterprise and on PCs

What happened

Nvidia launched new AI models that anyone can download and adjust. The big one, Nemotron 3.5 Super VL, is built for business AI agents. An agent is an AI program that can do tasks on its own. Nvidia also launched two smaller Bolt models that can run on a computer without sending data to the cloud.

Key numbers and facts

• Nemotron 3.5 Super VL has 120 billion parameters (the settings a model learns during training).
• Nvidia says Super VL can match the accuracy of the larger Nemotron 3 Ultra on some tasks, while using fewer parameters.
• Nemotron 3.5 Bolt 74B-A9B 24GB has 74 billion parameters and is built for local agent tasks.
• Bolt E4B is the smallest model and is aimed at chat-style agents. It arrives later this month.
• Super VL and Bolt 74B-A9B are available now.

Some background

Open-weight models let companies download the trained model and run it on their own computers or data centers. Nvidia says Super VL can be trained further on a company's own data. Nvidia also says it beats some rival models from Anthropic and OpenAI on certain agent tasks, such as document work, accounting, and insurance claims. Companies such as EY, Ramp, CrowdStrike, Perplexity, and Salesforce are named as users.

Why it matters

Many companies want AI they can run and control themselves, instead of relying only on outside services. Nvidia is also trying to sell more PCs with its chips by offering AI that runs on the device itself. This may help users keep private information, like health and money details, off the cloud. Still, these are Nvidia's own claims, and independent tests will show how well the models really work.

What to watch

• When Bolt E4B becomes available later this month.
• Whether independent tests confirm Nvidia's speed and accuracy claims.
• How well the new RTX Spark and DGX Spark PCs sell through partners like Microsoft and Dell.
• Whether companies adopt Super VL on their own data centers.