Nvidia’s most interesting move here isn’t a benchmark. It’s the argument that you should own your agent. The company launched new open-weight models built for always-on agents and edge computing, and the pitch is as much about control as capability.
The headline model is Nemotron 3.5 Super VL, a 120-billion-parameter model for enterprise agents. It uses fewer parameters than Nemotron 3 Ultra, but Nvidia says it matches that model’s accuracy on instruction-following and tool use. Nvidia also claims that, once post-trained on specific agentic tasks, Super VL finishes work faster than models from Anthropic and OpenAI, including Opus 5 and GPT 6.1 Sol, Luna and Astra. The task list covers document intelligence, financial accounting, scientific experimentation, medical insurance claims and security attack resilience.
Vendor benchmark claims are cheap, so the customer list matters more. Nvidia says Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce are using Super VL. That’s a credible set of names, though ‘using’ can mean anything from production deployment to a pilot.
At the small end, Nvidia released two Bolt models meant to run locally on new AI PCs with the RTX Spark chip and on DGX Spark workstations. Partners including Microsoft and Dell are shipping those machines now. Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter model for local agentic work, tuned to balance speed with reasoning for multistep tasks, tool calling and web browsing. The smaller Bolt E4B targets conversational agents.
Availability is straightforward. Super VL and Bolt 74B-A9B are available now, and Bolt E4B arrives later this month.
Why does this matter? Nvidia has a real problem. The frontier labs that buy its GPUs, and that Nvidia often invests in, are spreading their compute across AWS Trainium, Cerebras and their own custom silicon. Open models are Nvidia’s hedge. If enterprises run their own models in their own data centers, they need hardware, and Nvidia wants that hardware to be its own. Kari Briski, Nvidia’s vice president of Open Models and Software, made the case plainly: organizations can post-train Super VL on their own data, deploy it on their own infrastructure and own what they build.
This isn’t Nvidia’s first attempt. The Nemotron family already spans multimodal, reasoning, coding, math and enterprise automation. What’s new is the tiering. Nvidia is offering a model for each hardware class, from the data center down to the desk, which is the kind of packaging that Meta’s Llama and Mistral’s open releases have never fully matched on the hardware side.
The PC half is the riskier bet. Consumer AI assistants launch constantly, and many people won’t want an agent with access to their family details, health records and finances. Local execution is a sensible answer to that worry, since Briski says users can delegate everyday projects without sending workloads to the cloud. But it’s also a way to create demand for new Nvidia-powered PCs, and that’s a harder sell than a free model download.
So is this noise? Not entirely. The enterprise model is a competent answer to a real shift toward self-hosted agents. The PC models are more speculative. Watch whether developers actually adopt Bolt, because that will show whether local agents are a product category or just a hardware pitch.
