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

Nvidia on Thursday launched new open-weight models for always-on agents and edge computing, Julie Bort wrote in TechCrunch's report. The enterprise model is Nemotron 3.5 Super VL, a 120-billion-parameter model. It uses fewer parameters than Nemotron 3 Ultra, and Nvidia says it can deliver similar accuracy in instruction-following and tool-use tasks. Two lightweight Bolt models are intended to run locally on AI PCs with the Nvidia RTX Spark chip and on DGX Spark workstations.

Microsoft and Dell are releasing those PCs now for consumers, programmers, and enterprise knowledge workers. This piece covers the three models, the companies named as Super VL users, and the split between a data-center weight and a PC weight. It is for teams that want open models on their own Nvidia infrastructure, and for readers who want an assistant that keeps details about families, health, and finances off the cloud.

Nvidia launches new open models to run agents: the announcement

Kari Briski, Nvidia's vice president of Open Models and Software, wrote: "Any organization that wants to run agents on Nvidia infrastructure now has an open, optimized model for their tier without having to choose between capability and deployability." These are not Nvidia's first open-weight models aimed at proprietary frontier labs. Those labs are both Nvidia customers and companies Nvidia has invested in. The Nemotron family already covers multimodal artificial intelligence, reasoning, coding, math, and enterprise automation. The new weights add always-on agents and edge computing to that set.

Those frontier labs are also looking beyond Nvidia, to AWS's Trainium, Cerebras, and home-grown chips. Nvidia is leaning into enterprises that want to run their own open models in their own data centers. Briski wrote, "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." The two Bolt models are the PC half of Thursday's announcement, Nvidia's push into consumer assistants that run on the device.

What actually changed with Nvidia launches new open models to run agents

Nvidia claims Super VL completes tasks faster than competing models from Anthropic and OpenAI, named as Opus 5, GPT 6.1 Sol, Luna, and Astra, when post-trained on specific agentic tasks. Those tasks are document intelligence, financial accounting, scientific experimentation, medical insurance claims, and security attack resilience. Speed claims of this sort are standard whenever a major tech company announces a new model. Nvidia backed this one with heavy-hitter companies it says are already using Super VL: Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce.

The other new weights are the two lightweight Bolt models for the edge. Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter model for local agentic tasks, meant to balance speed with reasoning. It is intended to handle multistep reasoning, tool calling, and web browsing on the PC. The smallest model, Bolt E4B, is aimed at conversational agents. Beside them, Super VL is the 120-billion-parameter enterprise model, using fewer parameters than Nemotron 3 Ultra. Nvidia aimed the Bolts at local PCs rather than the data-center tier.

Who should care about Nvidia launches new open models to run agents

The enterprise audience is any organization that wants agents on Nvidia infrastructure and an open model for that tier. Briski wrote that buyers should not have to choose between capability and deployability. The Super VL workloads Nvidia listed are document intelligence, financial accounting, scientific experimentation, medical insurance claims, and security attack resilience. The companies it named as users are Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce. Those teams can post-train on their own data and deploy on their own infrastructure.

The PC audience is consumers, programmers, and enterprise knowledge workers. Microsoft and Dell are releasing the machines now: AI PCs on the Nvidia RTX Spark chip, and workstations running the DGX Spark. Nvidia hopes those computers, powered by its chips and running models locally, become a market of their own. Briski wrote that PC users "can delegate everyday projects ... without sending those workloads to the cloud." That is aimed at people who do not want agents reading about their families, health, and finances.

How to try Nvidia launches new open models to run agents

Super VL and Bolt 74B-A9B are available now, Nvidia says, and Bolt E4B comes later this month. The announcement states that schedule and does not state a dollar price. Enterprise teams are pointed at Nemotron 3.5 Super VL on Nvidia infrastructure: post-train it on their own data, deploy it on their own infrastructure, and own what they build. The local route is hardware already going out through Microsoft and Dell, an RTX Spark AI PC or a DGX Spark workstation.

Which model fits depends on the task. Super VL is the 120-billion-parameter enterprise model, with fewer parameters than Nemotron 3 Ultra, for instruction-following and tool-use after post-training on document intelligence, financial accounting, scientific experimentation, medical insurance claims, or security attack resilience. Nemotron 3.5 Bolt 74B-A9B 24GB, with 74 billion parameters, is the local model for multistep reasoning, tool calling, and web browsing. Bolt E4B, still due later this month, is the conversational agent. Super VL and the larger Bolt are the models available now.

What to watch after Nvidia launches new open models to run agents

Bolt E4B is the model Nvidia has not made available yet. Nvidia says it comes later this month, while Super VL and Nemotron 3.5 Bolt 74B-A9B 24GB are available now. Nvidia says post-trained Super VL finishes document intelligence, financial accounting, scientific experimentation, medical insurance claims, and security attack resilience faster than Opus 5, GPT 6.1 Sol, Luna, and Astra, and it names Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce as users. Published scores and timings are not in the announcement.

Other chips remain the open question here. Proprietary labs that are Nvidia customers and Nvidia investments are looking at AWS's Trainium, Cerebras, and home-grown chips. The launch answers with weights the enterprise can keep. Briski wrote, "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." On PCs, Nvidia hopes RTX Spark machines and DGX Spark workstations from Microsoft and Dell lead users to delegate everyday projects without sending those workloads to the cloud.