Nvidia’s latest AI release has less to do with selling GPUs directly and more to do with giving people a reason to keep buying them. The company has released a new set of open-weight Nemotron models aimed at two very different settings: large enterprise systems running agents around the clock, and AI PCs that run assistants entirely on local hardware.
The larger model, Nemotron 3.5 Super VL, has 120 billion parameters and is built for enterprise agent work. Nvidia says it matches the instruction-following and tool-use accuracy of its larger Nemotron 3 Ultra while using fewer parameters, which would make it cheaper and easier to deploy. The company also claims that, once post-trained for specific agentic jobs, Super VL finishes tasks faster than Anthropic’s Opus 5 and several OpenAI models. Those jobs include document processing, financial accounting, scientific experimentation, medical insurance claims and resilience against security attacks.
That condition matters. A model tuned for narrow workflows beating general-purpose frontier systems on those same workflows is a much smaller claim than broad superiority, and it has not been independently verified. Nvidia does point to a customer list that includes Salesforce, CrowdStrike, Perplexity, EY, Ramp and Skydance, which suggests the model is at least being used in real deployments.
At the smaller end are two Bolt models designed for machines built on Nvidia’s RTX Spark chip and DGX Spark workstations, a class of AI PCs now reaching stores through partners such as Microsoft and Dell. Nemotron 3.5 Bolt 74B-A9B, a 74-billion-parameter model, targets local agent tasks like multistep reasoning, tool calling and web browsing. The smaller Bolt E4B is aimed at conversational assistants. Super VL and Bolt 74B-A9B are available now, and Nvidia says E4B will follow later this month.
The strategy behind both releases is clear. OpenAI, Anthropic and other frontier labs are among Nvidia’s biggest customers and investments, yet they increasingly spread their compute across alternatives such as AWS Trainium, Cerebras and in-house chips. Open models that companies can fine-tune on their own data and run in their own data centres help keep Nvidia hardware at the centre of enterprise AI, whichever lab is ahead.
The PC push relies on a different selling point: privacy. Agents that handle finances, health records or family schedules are a much easier sell when that data stays on the device and never reaches the cloud. Whether consumers will want always-on assistants at all, local or not, is still unclear, and Nvidia is now betting a new hardware category on the answer.
