Nvidia's New Open Models Let You Run AI Agents Locally

According to TechCrunch, Nvidia has expanded its Nemotron family with open-weight models aimed squarely at autonomous agents. The release divides into two fronts: a massive 120-billion parameter model for corporate data centers and lightweight versions built to run directly on local hardware.

What are the Nemotron 3.5 Super VL and Bolt models?

The enterprise flagship is the Nemotron 3.5 Super VL, a 120-billion parameter model optimized for document intelligence, accounting, scientific research, and medical insurance claims. Despite having fewer parameters than older models, Nvidia claims it matches or beats competitors like Anthropic and OpenAI in tool use.

For local use, Nvidia released two Bolt models designed for AI PCs powered by the RTX Spark chip and workstations running DGX Spark. The Bolt 74B-A9B handles multistep reasoning and web browsing, while the smaller Bolt E4B focuses on conversational tasks.

How to use local AI agents in practice

Running these models locally changes how knowledge workers and developers handle sensitive projects on Monday morning. Instead of sending raw financial spreadsheets, medical records, or proprietary code to cloud APIs, your machine processes the workload on-device.

• Check your hardware: Ensure your workstation or new AI PC features supported Nvidia chips, such as the RTX Spark or DGX Spark architecture.
• Download or deploy the weights: Enterprise teams can pull the Super VL model into private data centers, while individual developers can access the Bolt models for local setups.
• Post-train on internal data: Customize the open weights using your company's proprietary documents without exposing data to third-party cloud providers.
• Execute multi-step tasks: Delegate browser automation, document analysis, and code generation to local agents that operate entirely offline.

Is it worth it for your workflow?

If your team handles strict compliance, financial records, or confidential client data, migrating to local open models solves the privacy roadblock of cloud-based assistants. You own what you build and keep workloads off external servers.

However, the catch lies in hardware requirements. Running models like the Bolt 74B locally demands dedicated processing power and memory (such as 24GB setups), making it impractical for older laptops. For heavy data privacy, it is a game-changer; for casual users, the hardware upgrade cost requires careful thought.