NVIDIA has introduced new open-weight artificial intelligence models designed to power enterprise AI agents and run AI tasks locally on compatible computers. The launch expands the company’s Nemotron family and reflects a growing push to make AI more customizable, efficient and accessible beyond cloud-based services.
The new models include Nemotron 3.5 Super VL, designed for complex enterprise workflows, and two smaller models in the Nemotron 3.5 Bolt family that target local AI applications on PCs and workstations.
The announcement matters because businesses increasingly want AI agents that can perform multi-step tasks, work with company data and use software tools. At the same time, consumers and professionals are exploring AI assistants that can operate directly on their computers without sending every task to a remote server.
Here is a closer look at NVIDIA’s new models, how they work and what they could mean for businesses and everyday computer users.
What Are NVIDIA’s New Nemotron 3.5 AI Models?
NVIDIA’s latest release includes three open-weight models intended for different AI workloads.
| Model | Main purpose | Key details |
|---|---|---|
| Nemotron 3.5 Super VL | Enterprise AI agents | 120 billion parameters; designed for complex workflows and visual understanding |
| Nemotron 3.5 Bolt 74B-A9B 24GB | Local AI agents | Designed for multi-step tasks, tool use and web browsing on compatible local hardware |
| Nemotron 3.5 Bolt E4B 4GB | Lightweight AI assistants | Designed for conversational and embedded applications with a smaller memory footprint |
According to NVIDIA, Super VL and Bolt 74B-A9B were available at launch, while Bolt E4B was scheduled to arrive on October 23, 2026.
The models are designed to cover different deployment needs, from enterprise data centres to local systems and smaller devices.
Rather than offering one model for every situation, NVIDIA is providing options that businesses and developers can select according to their hardware, privacy, speed and task requirements.
1. Nemotron 3.5 Super VL: Built for Enterprise AI Agents
Nemotron 3.5 Super VL is a 120-billion-parameter open-weight model designed for enterprise AI agents that handle complex tasks.
AI agents are systems that can do more than answer a single question. They can plan a sequence of actions, use tools, gather information and complete multi-step workflows with varying degrees of human supervision.
Super VL is designed to support activities such as:
- Following detailed instructions.
- Using software tools to complete tasks.
- Understanding visual information.
- Handling multi-step business workflows.
- Working with specialised enterprise data.
NVIDIA says Super VL can deliver performance comparable to its larger Nemotron 3 Ultra model on certain agentic tasks while using roughly one-fifth of the parameters.
The company also claims that Super VL can complete tasks up to five times faster than the next-best model in its comparison category. This is a vendor-reported benchmark claim, and actual performance can vary depending on the task, hardware and evaluation method.
One important feature is customisation. Organisations can post-train the model using their own data and deploy it on their own infrastructure.
This gives companies more control over how their AI systems are configured and where their workloads are processed.
2. Nemotron 3.5 Bolt: Bringing AI Agents to Local PCs
The Bolt models focus on a different opportunity: running AI agents on compatible local hardware rather than relying entirely on cloud services.
Nemotron 3.5 Bolt 74B-A9B 24GB is designed for agentic tasks that require reasoning, tool use and multiple steps.
NVIDIA says the model uses a native two-bit quantisation approach to fit within a 24GB memory footprint. Quantisation reduces the precision used to represent model weights, helping reduce memory requirements and make deployment on smaller systems more practical.
The model is intended to support tasks such as:
- Researching information on the web.
- Writing and editing content.
- Debugging code.
- Using tools through compatible agent frameworks.
- Completing everyday multi-step projects.
The benefit of local execution is that supported tasks can be processed on the user’s own computer, subject to hardware capability and software configuration.
This may reduce dependence on cloud services and give users more control over their data. However, local processing does not automatically guarantee privacy or security; those also depend on the software, permissions and system configuration.
3. What Is Nemotron 3.5 Bolt E4B?
Nemotron 3.5 Bolt E4B is a smaller model designed for conversational agents and applications with limited memory.
NVIDIA lists a four-billion-parameter model in the four-gigabyte category, targeting systems where memory efficiency is especially important.
Potential applications include:
- Voice assistants.
- Conversational interfaces.
- Interactive characters.
- Embedded AI applications.
- Lightweight assistants running on supported hardware.
Smaller models can be useful when low latency, lower memory use and local execution matter more than handling the most difficult reasoning tasks.
However, a smaller model will not necessarily perform as well as a larger one on every task. Developers must balance model capability, response speed, memory requirements and the complexity of the application.
NVIDIA said Bolt E4B was expected to become available on October 23, 2026.
Why Is NVIDIA Focusing on Open-Weight AI Models?
Open-weight models allow developers to download and run model weights under the terms of the relevant licence. Depending on the licence, users may also be able to customise or redistribute the model.
This differs from many closed AI services, where users access a model through a hosted product or API without receiving its weights.
NVIDIA’s strategy addresses several needs.
Greater control over deployment
Businesses may prefer to run AI systems inside their own data centres or controlled environments rather than sending sensitive workloads to external services.
Customisation for business tasks
Organisations can adapt suitable open-weight models to internal terminology, workflows and specialised requirements, subject to licensing and technical constraints.
More deployment options
A company may use a larger model for complex tasks and smaller models for simpler workloads, choosing the hardware and deployment environment that best fits each use case.
Reduced dependence on one hosted service
Running a model on local infrastructure can give an organisation more control over availability and operational decisions. However, it still needs to manage hardware, software, security, updates and performance.
Open weights do not mean that every part of a model is open source. Training data, full training processes and other components may remain unavailable.
How Could NVIDIA’s New Models Help Businesses?
Enterprise AI agents are becoming a focus for organisations seeking to automate repetitive work and help employees complete complex tasks.
NVIDIA’s new models could be useful in several areas.
Customer support
An AI agent could search approved knowledge bases, summarise customer information and help staff prepare responses. Human review may still be necessary for sensitive decisions.
Software development
Developers could use compatible agents to research code, generate tests, identify bugs and suggest fixes.
Document processing
Businesses could build systems that extract information from documents, compare records and prepare summaries.
Research and analysis
An agent could gather information from approved sources, organise findings and create a draft report for review.
Internal business operations
Companies could automate parts of reporting, data preparation and workflow coordination, provided that the agents have suitable permissions and safeguards.
These are potential applications, not guaranteed results from adopting a particular model. Real-world performance depends on the quality of the implementation, data access, tools, evaluation and human oversight.
Can Local AI Improve Privacy?
Running AI locally can reduce the need to transmit some information to a cloud provider.
For example, a user might process a document on a compatible computer instead of uploading it to a remote service. This can be useful when working with confidential business information or personal files.
Nevertheless, local AI is not automatically private or secure.
Applications may still connect to the internet, use external tools or transmit information depending on their settings. Local software can also have vulnerabilities, and agents that can access files or execute actions may create additional risks.
Users and businesses should therefore review permissions, network access, data storage, model licences and security controls before deploying an AI agent.
For enterprise systems, access controls and monitoring remain important even when the model runs on company-owned infrastructure.
What Hardware Is Needed to Run NVIDIA’s Bolt Models?
Hardware requirements depend on the specific model, its quantisation, the software used and the tasks being performed.
NVIDIA describes Bolt 74B-A9B 24GB as designed for compatible NVIDIA RTX and DGX Spark systems with a 24GB memory footprint.
The smaller Bolt E4B is intended for compatible platforms including GeForce RTX, Jetson and RTX/DGX Spark systems.
However, a model’s stated memory footprint does not mean it will run equally well on every device with the same amount of memory. Performance also depends on available system resources, software optimisation and the workload.
Before installing a local AI model, users should check NVIDIA’s official documentation for supported hardware, software dependencies, licence terms and deployment instructions.
How Does This Release Affect the AI Market?
NVIDIA’s announcement highlights two related trends in artificial intelligence.
The first is the growing interest in enterprise AI agents that can complete long-running workflows rather than simply respond to prompts.
The second is the effort to run more AI locally on personal computers and workstations.
These trends could create demand for a wider range of models and hardware configurations. Some workloads may be best handled by large cloud-based models, while others may be suitable for smaller models running locally.
NVIDIA’s position is also commercially significant because it develops the chips and software used by many AI systems. Encouraging local and enterprise AI deployments can support demand for its hardware and computing platforms.
However, the launch does not establish that NVIDIA’s models will outperform every competing model or become the preferred choice for all businesses. Developers will continue comparing models based on accuracy, speed, cost, licensing and hardware compatibility.
What Should Developers Check Before Using the Models?
Developers interested in Nemotron 3.5 should evaluate the models against their own tasks rather than relying only on headline benchmarks.
Important considerations include:
- Accuracy: Does the model complete the required tasks correctly?
- Latency: How quickly does it respond on the intended hardware?
- Memory use: Can it run alongside the other software the application needs?
- Tool access: Can it safely use the required APIs, browsers or software tools?
- Privacy: What information is stored or transmitted by the complete application?
- Security: Are permissions, monitoring and human approval controls appropriate?
- Licensing: Does the model licence allow the intended commercial or personal use?
- Total cost: How do hardware, deployment, maintenance and energy costs compare with hosted alternatives?
These checks can help developers determine whether an open-weight model is suitable for their requirements.
Conclusion
NVIDIA’s Nemotron 3.5 release expands its open-weight AI model family across enterprise systems and local computing devices.
Super VL targets complex enterprise AI agents, while Bolt 74B-A9B and Bolt E4B focus on local agentic and conversational applications with different hardware requirements.
The announcement reflects a broader shift towards AI systems that can use tools, perform multi-step tasks and operate in environments chosen by the user or organisation.
For businesses, the main opportunity is greater control over customisation and deployment. For individual users, local AI may offer more flexibility and reduce the need to send every task to the cloud.
The best choice will depend on the task, hardware, privacy requirements, cost and performance. As AI agents become more capable, careful testing and security controls will be just as important as the models themselves.
Frequently Asked Questions (FAQs)
1. What new AI models did NVIDIA launch in October 2026?
NVIDIA introduced Nemotron 3.5 Super VL, Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB.
2. What is Nemotron 3.5 Super VL used for?
It is designed for enterprise AI agents that handle complex instructions, visual understanding, tool use and multi-step business workflows.
3. What is Nemotron 3.5 Bolt?
Bolt is a family of models designed to bring agentic and conversational AI capabilities to compatible local computers and devices.
4. Can NVIDIA’s new models run without cloud processing?
Compatible deployments can run locally on supported hardware, allowing certain tasks to be completed without sending those workloads to the cloud. Some applications may still require internet access or external services.
5. Are NVIDIA’s Nemotron models open source?
NVIDIA describes these models as open-weight. Users should check each model’s licence and accompanying materials because open weights do not necessarily mean that all training data and processes are open source.
6. What is the difference between Super VL and Bolt?
Super VL targets complex enterprise workflows, while Bolt models are designed for local agentic or conversational applications with smaller memory requirements.
7. Which devices support Nemotron 3.5 Bolt?
NVIDIA lists compatible RTX and DGX Spark systems for Bolt 74B-A9B, and platforms including GeForce RTX, Jetson and RTX/DGX Spark for Bolt E4B. Actual compatibility depends on hardware and software configuration.
8. When did the new NVIDIA models become available?
NVIDIA said Super VL and Bolt 74B-A9B were available at launch. Bolt E4B was scheduled for October 23, 2026.
