Nemotron 3.5 Super VL: 120B Parameter Model Powers Efficient, Customizable Agentic AI for Enterprise

The short version

• NVIDIA released Nemotron 3.5 Super VL and two Bolt variants to support agent deployments from data centers to desktops.
• The Super VL model provides frontier-level accuracy for multimodal tasks while being five times faster than the next best open model.
• Edge models allow local processing of agentic tasks without cloud dependency, balancing performance with limited hardware resources.

The details

NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter model designed for specialized always-on agents. According to the company's blog, this model delivers the fastest task completion and highest accuracy in its class. It achieves performance comparable to NVIDIA’s largest reasoning model but uses roughly one-fifth of the parameters. The release aims to help enterprise agents complete high-volume tasks that span days or weeks more efficiently than previous options.

Two new edge models accompany the launch: Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB. The 74B-A9B variant uses native 2-bit quantization to fit within 24GB of memory, allowing it to run locally on NVIDIA RTX or DGX Spark systems. It is optimized to balance intelligence and speed for local agentic workloads. The smaller E4B 4GB model targets physical and conversational agents on GeForce RTX and Jetson hardware, supporting fast local voice assistants and interactive characters.

Super VL and Bolt 74B-A9B are available immediately, while Bolt E4B will launch on October 23. These three models join existing Nemotron family members, including Lightning for high-throughput workloads and Ultra for complex reasoning tasks. This expansion provides organizations with an open, optimized model for every tier of deployment, from large enterprise servers down to local edge devices, removing the need to choose between capability and deployability.

The context

The Nemotron Coalition collaborated on building Super VL, which demonstrates leading accuracy in instruction following, tool use, and visual understanding. Unlike closed frontier models, organizations can post-train Super VL on their own data and deploy it on their own infrastructure. This allows companies to oversee what they build and customize the model for specific policies and workflows. Dozens of organizations are already customizing the model for specialized tasks across various industries.

Scale AI post-trained the model for Skydance’s vendor credit finance agent and EY’s M&A due diligence agent. AfterQuery customized it for Hebbia’s credit agreement analysis, while AI21 and Alice focused on cybersecurity attack resilience. Mercor trained the model on coding benchmarks for agents that plan and debug code. Other users include Ramp for financial document intelligence, Dream for vulnerability analysis, Lila Sciences for scientific discovery, and R1 for medical claims analysis. TrainLoop post-trained the model with dermatology data from Nolla, which just became the first company approved for AI to write initial patient prescriptions.

What to watch

• Monitor the release of Nemotron 3.5 Bolt E4B 4GB on October 23 for edge device compatibility updates.
• Observe how post-training results from early adopters like Scale AI and AfterQuery impact specialized workflow accuracy.
• Track performance benchmarks on PinchBench and BrowseComp as new open models enter the agentic landscape.
• Watch for further integrations with local agent harnesses that enable web research and code debugging without cloud transmission.

What it means for organisations

Teams usually need to assess whether local processing aligns with their data privacy requirements before adopting edge models. Check whether current hardware specifications support the memory constraints of the new Bolt variants. Verify that post-training procedures comply with internal governance policies for customized AI agents. Ensure that stakeholders understand the trade-offs between cloud-based reasoning and local execution speed. Coordinate with security teams to evaluate risks associated with deploying open models on proprietary data.

Background: Agentic AI

Agentic AI refers to systems that plan and execute complex sequences of actions to achieve a goal, rather than just answering questions. Unlike standard chatbots, these agents can interact with other software, browse the web, and run commands. This autonomy increases efficiency but also expands the potential for security breaches if the agent’s permissions are not strictly limited.

Read the full guide: Agentic AI Explained: How Autonomous Agents Change Security

What to do and how to stay safe: NVIDIA

• Audit existing agent deployments to identify workloads that could benefit from local processing to reduce latency and cloud costs.
• Review hardware inventory to determine if current RTX or DGX Spark systems meet the memory requirements for the new Bolt models.
• Establish clear guidelines for post-training open models on proprietary data to ensure compliance with data protection regulations.
• Monitor vendor documentation for updates on model compatibility and performance benchmarks before integrating into production environments.

Step-by-step guide: Secure Home Routers for Small Business Networks

General security guidance from the TechSec Headlines newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is Nemotron 3.5 Super VL?

It is an open 120-billion-parameter model that delivers high accuracy and fast task completion for specialized always-on agents, using one-fifth the parameters of larger reasoning models.

When are the new models available?

Nemotron 3.5 Super VL and Bolt 74B-A9B 24GB are available now, while Bolt E4B 4GB will be released on October 23.

Can organizations customize these models?

Yes, organizations can post-train Super VL on their own data, tools, and policies to improve accuracy for specialized tasks and deploy it on their own infrastructure.