Nvidia has introduced open-weight models for enterprise agents and on-device PC workflows. Nemotron 3.5 Super VL is a 120-billion-parameter model focused on enterprise tasks such as document analysis, financial accounting, and health insurance claims.
For PCs, Nvidia has introduced the Bolt family, designed to run locally on AI PCs powered by Nvidia RTX Spark chips and DGX Spark workstations. The Nemotron 3.5 Bolt 74B-A9B 24GB supports multi-step tasks, tool use, and web browsing, while Bolt E4B focuses on conversational agents.
Super VL and Bolt 74B-A9B are now available, while Bolt E4B is expected later this month.
Nvidia’s New Models
Nemotron 3.5 Super VL is a 120-billion-parameter model for enterprise agents. Nvidia says it delivers instruction-following and tool-use accuracy close to Nemotron 3 Ultra despite using fewer parameters.
For smaller machines, there is Bolt 74B-A9B 24GB for on-device agent tasks and Bolt E4B, which focuses on conversational agents. Super VL and Bolt 74B-A9B are currently available, while Bolt E4B will arrive later this month.
When Agents Need to Work with Private Data Without Sending It to the Cloud
When agents help manage family, health, financial, or internal business data, privacy concerns are difficult to avoid—especially when that data is sent to the cloud for processing.
Nvidia’s approach is to run agents on an organization’s own infrastructure while allowing Super VL to be post-trained with its own data. PC users can also delegate routine tasks to models running locally, without sending their work to the cloud.
Where Super VL and Bolt Fit in the Nemotron Family
The Nemotron family covers tasks ranging from multimodal processing, reasoning, coding, and mathematics to enterprise automation. Super VL is the large model for enterprise agent tasks, with 120 billion parameters and a focus on instruction-following and tool use.
Bolt is the lightweight line for running locally on both AI PCs powered by RTX Spark chips and DGX Spark workstations. Bolt 74B-A9B 24GB supports multi-step reasoning, tool use, and web browsing, while Bolt E4B focuses on conversational agent tasks.
From Nemotron 3 Ultra to Super VL
Super VL is a model for enterprise agents with 120 billion parameters—fewer than Nemotron 3 Ultra. However, Nvidia claims that it still delivers comparable accuracy on instruction-following and tool-use tasks.
| Factor | Nemotron 3 Ultra | Nemotron 3.5 Super VL |
|---|---|---|
| Number of parameters | More than 120 billion | 120 billion |
| Task profile | Large-scale model | Enterprise agents |
| Instruction-following and tool use | Benchmark reference | Nvidia claims accuracy close to Nemotron 3 Ultra |
Capabilities Demonstrated Through Real-World Tasks
For document-heavy work that requires understanding large amounts of information, the model can be used for document intelligence to help read and organize data systematically.
In finance, the model supports accounting tasks and health insurance claim reviews. In research, it can be used for scientific experiments.
For Nemotron 3.5 Bolt 74B-A9B, the key strengths are multi-step reasoning, tool use, and web browsing. This makes it suitable for agents that need to perform continuous, multi-stage tasks on internal organizational infrastructure or supported PCs.
Super VL Compared with Models from Anthropic and OpenAI
Nvidia says Super VL completes tasks faster than Opus 5, GPT 6.1 Sol, Luna, and Astra after being post-trained for specialized agent tasks such as document processing, accounting, and scientific experimentation. These are claims from Nvidia, not results from independent testing.
| Factor | Super VL | Opus 5 | GPT 6.1 Sol | Luna | Astra |
|---|---|---|---|---|---|
| Nvidia’s stated result | Faster after post-training for specialized tasks | Comparison model | Comparison model | Comparison model | Comparison model |
| Task profile | Specialized agent tasks | Specialized agent tasks | Specialized agent tasks | Specialized agent tasks | Specialized agent tasks |
Strengths and Limitations to Know Before Using Them
Pros
- +Can be customized with organizational data and deployed on the organization’s own infrastructure
- +Can run certain tasks locally, reducing the need to send data to the cloud
- +Offers multiple model tiers, from Super VL for enterprises to Bolt for PCs
Cons
- −Usage is tied to Nvidia’s infrastructure and hardware
- −Speed claims are based on Nvidia’s testing and specialized agentic tasks
- −Super VL and Bolt 74B-A9B are available now, but Bolt E4B will arrive later this month
Costs Beyond the Model Specifications
Using Bolt locally still requires acquiring a machine with an RTX Spark chip or a workstation running on DGX Spark. The cost therefore includes more than the model itself—it also covers hardware selection and maintenance.
Organizations must prepare their own data center infrastructure, as well as install and maintain the systems needed to run the models in production. Post-training with internal data also requires additional resources and processes.
The advantage is that organizations retain ownership of their data and what they build. However, the responsibility for deployment and maintenance shifts entirely to internal teams. PC users must likewise invest in hardware that supports local agent execution if they want to avoid sending tasks to the cloud.
Who Is This Approach For?
Nvidia’s solution is suited to organizations that want to run agents in their own data centers and post-train them with internal data while retaining ownership of what they build. Programmers and knowledge workers using RTX Spark PCs or DGX Spark workstations are also part of the target audience.
Made for
- Organizations that want to run agents on their own infrastructure
- Teams that want to control their data and post-train models themselves
- Programmers and knowledge workers using Nvidia PCs or workstations
Think twice
- Teams that are not yet ready to manage deployment and infrastructure themselves
Skip this one
- Users who do not want to manage hardware — consider using cloud-based models
What Nvidia Is Selling Alongside the New Models
This launch is not just about selling models; it also promotes a vision of running them on Nvidia infrastructure. Organizations can run the models in their own data centers, while the AI PC market gets machines that run agents through RTX Spark chips and DGX Spark.
The key question is how much locally run agents will help users control their data. Some tasks may not need to be sent entirely to the cloud, but managing the hardware and systems independently remains a significant challenge.