Nvidia Unveils Open AI Models for Enterprise Agents and Local PCs

Nvidia on Thursday introduced a new set of open-weight AI models aimed at two fast-growing markets: enterprise agent systems running in data centers and lightweight assistants that can run directly on personal computers. The launch adds fresh momentum to the company’s push to make its software stack central to both cloud and on-device AI.

The headline release, Nemotron 3.5 Super VL, is a 120-billion-parameter model Nvidia says is built for enterprise agents that need strong instruction-following, tool use and multimodal reasoning. At the same time, the company unveiled two smaller Bolt models designed for local execution on Nvidia-powered AI PCs and workstations, including systems built around the new RTX Spark and DGX Spark platforms.

The move matters because it gives businesses a new open model option for deploying AI agents on their own infrastructure while also signaling Nvidia’s bet that a privacy-conscious consumer market will emerge for assistants that work offline or with minimal cloud dependence.

What Nvidia announced

Nvidia’s announcement centers on a family of open-weight models built for different deployment tiers. The company is not simply releasing one model to compete with the leading proprietary AI systems; it is trying to cover the full spectrum from large enterprise deployments to local desktop use.

According to Nvidia, Nemotron 3.5 Super VL is intended for organizations that want agentic systems capable of handling complex work across documents, tools and multiple modalities. The company says the model can match the accuracy of its larger predecessor, Nemotron 3 Ultra, while using fewer parameters. On the smaller end, the Bolt models are meant to run locally on new AI PCs and workstations rather than in the cloud.

Why Nvidia is betting on open models

Nvidia’s timing reflects a broader shift in enterprise AI purchasing. Many companies want the power of frontier models without fully depending on closed APIs or sending sensitive data to outside cloud services. By offering open-weight models, Nvidia is trying to position itself as the infrastructure layer for organizations that want control over deployment, fine-tuning and data governance.

The company’s vice president of Open Models and Software, Kari Briski, said in a blog post that organizations can now choose an Nvidia model for their specific deployment tier without sacrificing either capability or practicality. In her framing, the appeal is not just model quality but the freedom to adapt and operate it internally.

Briski said organizations can post-train Super VL on their own data, run it on their own infrastructure and keep ownership of what they build, rather than relying on closed frontier systems.

That message is especially relevant as major AI labs increasingly diversify their hardware suppliers. Some are working with alternatives such as AWS’s Trainium or Cerebras, while others are designing their own chips. Nvidia’s open-model push helps it stay embedded in AI adoption even when its customers are also its rivals.

How does Super VL compare with rival models?

Nvidia says Super VL is faster than competing systems from Anthropic and OpenAI when the model is post-trained for specific agentic workflows. The tasks Nvidia highlighted include document intelligence, accounting, scientific experimentation, medical insurance claims and resilience against security attacks.

Those are high-value enterprise workloads, and they are also the kinds of tasks where reliability matters as much as raw benchmark performance. Nvidia is clearly aiming to persuade customers that agent systems can be practical business tools rather than experimental chat interfaces.

As with most model launches, Nvidia’s comparisons should be read as vendor claims rather than independent verdicts. But the company is backing up its pitch with a set of named customers that it says are already using Super VL, including Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce.

Why the enterprise angle matters

Enterprise buyers have become more comfortable with open-weight models because they can be tuned for specialized tasks, deployed behind company firewalls and integrated into existing security and compliance frameworks. For many CIOs and AI teams, that flexibility is the difference between experimenting with AI and actually rolling it out at scale.

Nvidia is trying to make that transition easier by bundling model capabilities with its hardware ecosystem. The company’s argument is straightforward: if an organization wants an agent system that is powerful enough for real work and still manageable on-premises, Nvidia wants to be the default provider.

What the Bolt models mean for local AI

The Bolt models are Nvidia’s bet on a second trend: consumer and professional AI assistants that run locally on PCs instead of sending every interaction to the cloud. That shift could appeal to users who want lower latency, more control over data and less dependence on online services.

The larger of the two, Nemotron 3.5 Bolt 74B-A9B 24GB, is designed to balance speed and reasoning for multistep tasks such as tool use and web browsing. The smaller Bolt E4B is targeted at conversational agents. Nvidia says the 74B model is already available, and the smaller model will arrive later this month.

These local models are meant to run on the company’s AI PCs and workstations powered by the new RTX Spark and DGX Spark chips. Nvidia says those machines are entering the market now through partners including Microsoft and Dell.

How local execution changes the user experience

Running an assistant locally changes more than where the computation happens. It can also change what kinds of tasks people are willing to delegate. A local model can potentially help with family records, health-related information, financial documents or other sensitive material without forcing users to upload those details to a cloud service.

That privacy argument is important because AI assistants are becoming more capable at managing everyday digital work, but that same capability raises trust concerns. Nvidia is trying to make the case that if consumers are hesitant to share deeply personal information with remote AI services, then the PC should become the new home for the assistant.

There is also a commercial motive. If Nvidia can create demand for AI PCs that rely on its chips, it expands beyond the data center and into the consumer hardware market. That gives the company another route to monetization as AI moves closer to the desktop.

How Nvidia is trying to stay ahead of the AI market shift

Nvidia’s broader strategy is to remain indispensable as the AI ecosystem fragments. OpenAI, Anthropic and other frontier-model companies remain hugely influential, but they are no longer the only centers of gravity in the market. Enterprises are increasingly asking whether they can run models themselves, customize them more deeply and reduce dependency on a single vendor.

That is where Nvidia’s open-weight family becomes strategically valuable. It lets the company compete not only on raw model quality but also on portability, deployability and ecosystem control. Instead of forcing customers into a cloud-only experience, Nvidia is offering them a stack that spans model, chip and deployment environment.

The company already has a broader Nemotron lineup covering multimodal AI, reasoning, coding, math and enterprise automation. The new releases extend that approach by splitting the market into two directions at once: heavier, business-focused systems for data centers and compact models for PCs.

How the enterprise and PC strategies fit together

At first glance, enterprise AI agents and local desktop assistants may seem like separate bets. In practice, Nvidia is treating them as complementary parts of the same market transition.

On one end, large companies want agentic software that can handle business processes, analysis and workflow automation. On the other, individual users and knowledge workers want responsive assistants that can operate without handing data to a remote provider. Both groups want control, and both need hardware optimized for the job.

That convergence creates an opening for Nvidia. If the same company can supply the models, the silicon and the reference systems for both categories, it can shape the architecture of AI adoption from the server rack to the desk.

• Enterprises gain an open model they can adapt and host internally.
• PC makers gain a reason to build AI-first systems around Nvidia chips.
• End users gain local assistants with lower cloud dependence.
• Nvidia gains a broader market for both software and hardware.

What this says about the broader AI race

The launch also reflects a changing competitive landscape. For years, the conversation in generative AI centered on which labs had the biggest and most capable proprietary models. Now the market is increasingly about deployment flexibility, cost, privacy and control.

That shift plays to Nvidia’s strengths. The company is not just a chip supplier; it is increasingly presenting itself as the platform on which different categories of AI can run. By releasing open models that are aligned with its infrastructure, Nvidia can stay relevant even as customers test alternatives to its hardware or build custom systems of their own.

It also helps Nvidia answer a growing concern among buyers: if AI is becoming essential, how do you avoid lock-in? Open-weight models are one answer, and Nvidia is making the case that its own versions can offer the same freedom without sacrificing performance.

What to watch next

The immediate questions are whether businesses adopt Super VL broadly, whether developers fine-tune it for specialized agentic workflows and whether the Bolt models can help spark demand for local AI PCs. Nvidia will also need to prove that local deployment is not just technically feasible but commercially useful for everyday users.

Another question is how competitive the models will look once independent benchmarks and real-world deployments emerge. Company claims about performance are common in AI launches, but the market increasingly wants proof from customers, developers and third-party evaluators.

Still, Nvidia’s announcement is notable for one reason above all: it shows the company expanding its AI strategy beyond the cloud into places where control, privacy and local execution are becoming just as important as scale.

In a market where AI vendors are increasingly competing on where models run as much as how well they perform, Nvidia’s latest release is a clear signal. The company wants to own both the enterprise agent stack and the future of local AI on PCs.