Nvidia brings open AI agents from servers to PCs

A smaller model for enterprise agents

Nemotron 3.5 Super VL is designed for enterprise AI agents. Nvidia says it has fewer parameters than Nemotron 3 Ultra but reaches comparable accuracy in instruction-following and tool use.

The company also claims that, after fine-tuning for specialized tasks, Super VL runs faster than competing models from Anthropic and OpenAI: Opus 5, GPT 6.1 Sol, Luna and Astra. The listed workloads include document analysis, financial accounting, scientific experiments, medical insurance claims processing and resistance to attacks.

Nvidia named Skydance, EY, Ramp, Crowdstrike, Perplexity and Salesforce as companies already using the model. Those references offer a useful signal of interest, but the announcement does not provide deployment details or results from those users.

Super VL is available now. Nvidia has not published the comparison results in the material described here, so its speed claim is difficult to assess from the announcement alone.

Open models as a platform bet

Super VL adds to Nvidia’s Nemotron family, which already spans multimodal AI, reasoning, coding, mathematics and enterprise automation. The company is positioning open models as a way for organizations to fine-tune systems on their own data and run them on their own infrastructure.

That pitch matters as fast-growing labs developing closed models increasingly turn to other chip and AI-system providers, including AWS’s Trainium and Cerebras. Some are also building their own chips. Nvidia’s open-model strategy gives it another reason for customers to keep using its infrastructure, even as they seek alternatives to closed systems.

Nvidia vice president of open models and software Kari Briski said organizations running agents on Nvidia infrastructure can choose an optimized open model for each level of work without giving up capability or deployment convenience. Briski also said customers can fine-tune Super VL on their data and retain control of the resulting systems.

I think the strategic case is clearer than the performance case so far: Nvidia is selling control over where models run, while its headline speed comparisons still need evidence that readers can inspect.

Bringing agents onto PCs

Nvidia’s two compact Bolt models are aimed at AI PCs with its RTX Spark chip and workstations with DGX Spark. Microsoft and Dell are among the partners bringing those computers to market.

The models target different local workloads:

Both products extend Nvidia’s enterprise argument to personal computing: running agents locally could let users delegate everyday tasks without sending them to the cloud. Nvidia argues that this may matter to people reluctant to share information about family, health, finances and other personal matters with online assistants.

Super VL and Bolt 74B-A9B are available now. Nvidia says Bolt E4B will arrive later this month.

What I’d want to know is how much of the promised privacy and convenience survives in real use: the announcement describes local models, but says little about what these agents can do reliably on a PC. If Nvidia can make local deployment useful beyond a hardware showcase, the PC becomes another place where its chips and open models reinforce each other.