Nvidia on Thursday expanded its Nemotron family with a set of open-weight models aimed squarely at always-on AI agents, splitting the release between a large enterprise model and two lightweight variants built to run locally on AI PCs and workstations. The move puts Nvidia's weight behind a simple proposition: organizations that want to run agents on their own infrastructure should be able to do so with an open, optimizable model at whichever tier of compute they already have.

A large model for enterprise agents

The headline release is Nemotron 3.5 Super VL, a 120-billion-parameter model intended for enterprise agent workloads. It is smaller than Nvidia's Nemotron 3 Ultra, but the company says it delivers comparable accuracy on instruction-following and tool-use tasks, the two capabilities that matter most when an agent has to translate a request into a sequence of reliable actions. For teams building document pipelines, financial workflows or security automation, that trade-off between parameter count and deployability is the central engineering question.

Nvidia claims Super VL completes agentic tasks faster than competing models from Anthropic and OpenAI, including Opus 5, GPT 6.1 Sol, Luna and Astra, when it is post-trained on specific tasks such as document intelligence, financial accounting, scientific experimentation, medical insurance claims and security attack resilience. Such comparisons are routine when a major vendor ships a model, but Nvidia paired the claim with a roster of organizations it says are using Super VL: Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce. Those names span media production, professional services, fintech, security, search and CRM, which gives the launch a broader commercial footprint than a typical model drop.

Bolt models bring agents to the desktop

At the other end of the range, Nvidia introduced two Bolt models designed to run locally rather than in a data centre. Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter model for local agentic work, balancing speed against reasoning so it can handle multistep reasoning, tool calling and web browsing. The smallest variant, Bolt E4B, targets conversational agents. The Bolt models are meant for the new AI PCs built around Nvidia's RTX Spark chip and for workstations running DGX Spark, hardware that partners including Microsoft and Dell are bringing to market now for consumers, programmers and enterprise knowledge workers.

Super VL and Bolt 74B-A9B are available now, while Bolt E4B is scheduled to arrive later this month. The staggered rollout lets Nvidia seed the larger enterprise model first while the smallest conversational variant follows.

Why open weights matter to Nvidia

Nvidia's pitch rests on control rather than raw capability. "Unlike closed frontier models, organizations can post-train Super VL on their own data, deploy it on their own infrastructure and own what they build," wrote Kari Briski, Nvidia's vice president of Open Models and Software, in the announcement. That framing speaks directly to enterprises that want agents tuned on proprietary data without shipping sensitive workloads to a third-party API. It also gives Nvidia a story for customers who are increasingly spreading their AI compute across alternatives such as AWS's Trainium, Cerebras and in-house silicon from the frontier labs themselves.

The local Bolt models extend the same logic to the desktop. By running agents on the machine rather than in the cloud, PC users can hand off everyday projects without sending those workloads off-device, a pitch aimed at privacy-conscious buyers who may hesitate to give a cloud assistant access to family, health and financial information. Whether that concern translates into demand for AI PCs is an open question, but it is the market Nvidia is trying to create for its chips.

An expanding but crowded portfolio

Super VL and the Bolt line join a Nemotron family that already covers multimodal AI, reasoning, coding, math and enterprise automation. The strategy is not new for Nvidia, which has released open-weight models before to compete with proprietary frontier labs that are simultaneously its customers and its investments. What is notable is the tiering: a single announcement now spans a 120-billion-parameter enterprise workhorse, a 74-billion-parameter local agent model and a small conversational model, each positioned for a different deployment target.

For developers, the practical question is how well Super VL holds up after post-training on domain-specific agent tasks and whether the Bolt models deliver usable tool-calling performance on RTX Spark and DGX Spark hardware. Nvidia's partner list and its availability dates give teams something concrete to test now, with the smallest Bolt model still to come before the end of the month.