NVIDIA has expanded its open-weight model portfolio with the Nemotron 3.5 family, a lineup split between large models aimed at enterprise AI agents and compact variants designed to run entirely on local hardware.

The flagship of the family is Nemotron 3.5 Super VL, a 120-billion-parameter model targeted at enterprise agent infrastructure. Alongside it, NVIDIA introduced two much smaller models under the "Bolt" name, built with local execution in mind.

Super VL: fewer parameters, similar accuracy

n1

Super VL carries fewer parameters than its predecessor, Nemotron 3 Ultra, yet NVIDIA says it achieves nearly the same accuracy on key agentic tasks such as instruction following and tool use. Kari Briski of NVIDIA said that organizations building agents on their own infrastructure now get an optimized open model without having to choose between capability and scalability.

According to the company, when Super VL is tuned for agentic scenarios such as document analysis, financial accounting, scientific experiments, medical insurance claims and resilience against security attacks, it completes work faster than models from Anthropic and OpenAI, including Opus 5, GPT 6.1 Sol, Luna and Astra. A customer list backing that claim includes well-known names such as Skydance, EY, Ramp, Crowdstrike, Perplexity and Salesforce.

Bolt models run on the desktop

n2

The Bolt models are designed to operate independently of the data center, running directly on next-generation AI PCs and DGX Spark-based workstations. They reach users through partners including Microsoft and Dell, and target developers, enterprise knowledge workers and end users alike.

With a new AI assistant appearing almost daily, privacy-conscious users may hesitate to hand over family data, health records or financial information to cloud services. NVIDIA emphasizes that with locally running models, everyday tasks and data can be handled directly on the user's own machine without ever touching the cloud.

The first model under the Bolt banner is Nemotron 3.5 Bolt 74B-A9B 24GB, a 74-billion-parameter model tuned for local agent work. It is intended to balance speed and reasoning in demanding processes such as multi-step reasoning, external tool calling and web browsing. The smallest member of the family, Bolt E4B, focuses on chat-oriented agents.

NVIDIA says Super VL and Bolt 74B-A9B are available now, while Bolt E4B is expected to reach users by the end of the month.