Nvidia launches open-weight models for enterprise and PC agents

Nvidia has introduced new open-weight AI models built for always-on agents in enterprise infrastructure and on local PCs. The move matters because Nvidia is pushing a private-agent path: companies and PC users can run models closer to their own data instead of relying only on cloud-based systems.

The launch, reported by TechCrunch, includes a large enterprise model called Nemotron 3.5 Super VL and two smaller Bolt models for AI PCs and workstations. Nvidia says the lineup is meant to give organizations different deployment options without forcing a trade-off between capability and where the model runs.

Three models aimed at different agent setups

The largest new model is Nemotron 3.5 Super VL, a 120-billion-parameter open-weight model designed for enterprise agents. Nvidia says it uses fewer parameters than Nemotron 3 Ultra but can reach similar accuracy on instruction-following and tool-use tasks.

The company also introduced two lighter Bolt models for local use on new AI PCs running Nvidia’s RTX Spark chip and on DGX Spark workstations. Those PCs are aimed at consumers, programmers and enterprise knowledge workers, and Nvidia says they are reaching the market through partners including Microsoft and Dell. Emonarc has also been tracking the broader RTX Spark Windows PC rollout.

The new models break down as follows:

• Nemotron 3.5 Super VL: a 120-billion-parameter enterprise model for agent workloads.
• Nemotron 3.5 Bolt 74B-A9B 24GB: a 74-billion-parameter local model for agentic tasks, including multistep reasoning, tool calling and web browsing.
• Bolt E4B: the smallest Bolt model, aimed at conversational agents.

Nvidia says Super VL and Bolt 74B-A9B are available now, while Bolt E4B is due later this month.

Nvidia’s pitch: keep agents closer to company data

Nvidia is positioning Super VL around the enterprise trend of running open models inside company-controlled infrastructure. Kari Briski, Nvidia’s vice president of Open Models and Software, wrote in the company’s announcement that organizations can post-train Super VL on their own data, deploy it on their own infrastructure and own what they build.

That is the key contrast with closed frontier models. Many teams already use cloud assistants such as ChatGPT and Claude for writing, research, coding and analysis, but enterprise agent projects often raise additional questions: where sensitive documents live, how tools are connected, and whether model behavior can be tuned around internal processes.

Nvidia’s broader Nemotron family already spans multimodal AI, reasoning, coding, math and enterprise automation, according to TechCrunch. This latest release narrows the focus toward agents that need to call tools, follow instructions and keep running inside enterprise environments.

The PC angle is about privacy and local delegation

The Bolt models point to a second market Nvidia wants to encourage: consumer and workplace AI assistants running locally on PCs. The company’s argument is that users may want agents to help with everyday projects without sending those workloads to the cloud.

That matters for tasks involving personal or business information, such as family details, health context, finances or internal work documents. TechCrunch notes that privacy-conscious users may hesitate to give cloud assistants broad access to that kind of information.

For freelancers, creators and small teams, the near-term takeaway is less about training huge models and more about the direction of the tooling. If local AI PCs become common, some agent workflows may shift from browser tabs and cloud dashboards to machines that can reason, browse and call tools from the desktop. Still, web-browsing agents face practical friction across the internet, as seen in wider pushback against AI agents used for booking and shopping.

Performance claims and enterprise proof points

Nvidia says Super VL can complete tasks faster than competing models from Anthropic and OpenAI when post-trained for specific agentic work. The examples named by the company include document intelligence, financial accounting, scientific experimentation, medical insurance claims and security attack resilience.

As with any vendor benchmark claim, buyers should treat the result as a starting point rather than a universal answer. The relevant question is whether the model performs well on the documents, tools and workflows a business actually uses.

Nvidia also pointed to companies it says are using Super VL, including Skydance, EY, Ramp, Crowdstrike, Perplexity and Salesforce. The customer list helps explain the enterprise emphasis: Nvidia is not only selling chips for AI labs, but also trying to make its own model stack attractive to organizations that want agents on Nvidia infrastructure.

There is a strategic backdrop here. TechCrunch notes that proprietary AI labs, which are both Nvidia customers and investments, are also looking at other chip and AI system options, including AWS Trainium, Cerebras and internally built chips. Open-weight models for enterprise and PC deployment give Nvidia another way to keep agent workloads tied to its hardware ecosystem.