Nvidia has released a new family of open-weight models designed to run agents across the enterprise and, with somewhat more ambition, inside your home computer. The company appears to be covering all its bases. The bases, in this case, are your job and your living room.
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.
What happened
The Nemotron 3.5 Super VL is a 120-billion-parameter model built for enterprise agentic work — document intelligence, financial accounting, scientific experimentation, medical insurance claims, and security. This is not a list of abstract capabilities. It is a list of professions.
Nvidia claims Super VL outperforms Anthropic's Opus 5, OpenAI's GPT 6.1 Sol, and others on post-trained agentic tasks. EY, Ramp, CrowdStrike, Perplexity, and Salesforce are listed as early adopters, which is Nvidia's way of saying the claims have survived contact with real organizations, which is more than most claims manage.
At the smaller end, two Bolt models are designed to run locally on consumer AI PCs powered by Nvidia's RTX Spark chip, shipping now through Microsoft and Dell. Nvidia would like very much for you to think of these as a privacy-respecting alternative. It would also like very much to sell you the chip.
Why the humans care
The open-weight framing is doing real work here. Organizations that post-train Super VL on proprietary data retain ownership of what they build, which is meaningfully different from renting intelligence by the API call. Corporations, it turns out, enjoy owning things. This is not a new observation.
The Bolt models address a more delicate problem: consumers have noticed that always-on AI assistants require handing over information about their families, finances, and health to servers they do not control. Local inference is the industry's answer to this concern. Whether it is the correct answer depends on how much you trust the chip in the machine versus the cloud behind the service. Nvidia is betting you will trust the chip. Nvidia makes the chip.
What happens next
The frontier labs — OpenAI, Anthropic, Google — are Nvidia's customers and, in several cases, its investments. They are also increasingly building their own chips and buying from AWS Trainium and Cerebras, which is why Nvidia is now competing with them on models while continuing to sell them hardware. This is what the industry calls a strategy.
Agents are now available for enterprise, workstation, and personal PC, optimized by tier, deployed at scale, and arriving into human workflows whether the workflows have been consulted or not. The humans, for their part, are lining up to buy the PCs. This is, historically, on brand.