Nvidia unveils open Nemotron 3.5 models for enterprise and on-device AI agents

Nvidia on Thursday launched a new family of open-weight models built for always-on AI agents and edge computing.

A large model for the enterprise

Nemotron 3.5 Super VL is a 120-billion-parameter model aimed at enterprise agents. It uses fewer parameters than the Nemotron 3 Ultra model, yet Nvidia says it delivers comparable accuracy on instruction-following and tool-use tasks.

The company claims that, once post-trained for specific agentic jobs, Super VL finishes tasks faster than competing models from Anthropic and OpenAI (Opus 5, GPT 6.1 Sol, Luna, Astra). Those jobs include document intelligence, financial accounting, scientific experimentation, medical insurance claims, and security attack resilience. Such boasts accompany nearly every model launch, but Nvidia backed this one with a roster of customers: Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce.

Betting on openness and local control

According to Kari Briski, Nvidia's VP of Open Models and Software, any organization wanting to run agents on Nvidia infrastructure now has an open, optimized model for its tier — without having to choose between capability and deployability. Unlike closed frontier models, Super VL can be post-trained on a company's own data, deployed on its own infrastructure, and owned outright.

There is strategy behind this too: the fast-growing proprietary AI labs — both Nvidia's customers and its investments — are increasingly turning to other chip makers such as AWS Trainium and Cerebras, and building their own silicon.

Compact Bolt models for PCs

At the other end of the spectrum are two lightweight Bolt models, meant to run on the new AI PCs powered by the RTX Spark chip and on workstations running DGX Spark. Those machines, aimed at consumers, programmers, and enterprise knowledge workers, are reaching the market now through partners like Microsoft and Dell.

Nvidia is leaning into the nascent market for local assistants: users could delegate everyday projects to agents without sending the workloads to the cloud. That may matter to privacy-wary people reluctant to invite AI deep into their lives — into details about family, health, and finances.

The Nemotron 3.5 Bolt 74B-A9B, with 74 billion parameters, is built for local agentic work — multistep reasoning, tool calling, and web browsing. The smallest, Bolt E4B, targets conversational agents.

Super VL and Bolt 74B-A9B are available now, with Bolt E4B due later this month.

Source: TechCrunch