NVIDIA Unveils New Open AI Models for Agents Across the Stack

NVIDIA just rolled out Nemotron 3.5 Super VL, a 120-billion-parameter open model, plus two new edge-oriented models: Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB. With these, NVIDIA says the Nemotron lineup now covers the spectrum—whether you’re running agents on enterprise servers or deploying to lean edge devices. Super VL and Bolt 74B-A9B are available immediately; Bolt E4B lands October 23.

What’s New: Details on Nemotron 3.5 Super VL and Bolt Models

Here’s what jumps out: Nemotron 3.5 Super VL is designed in collaboration with the Nemotron Coalition. It’s a 120B-parameter open-weight model targeting high-accuracy agentic tasks—instruction following, tool integration, multistep workflows, and reading visual documents. NVIDIA claims Super VL matches their largest reasoning model for accuracy but shrinks the size to a fifth, and on the Artificial Analysis Intelligence Index leaderboard, it delivers up to 5x the throughput versus peers.

The open format means you can post-train the model with your data and run it on your own stack.

On the edge, Bolt 74B-A9B 24GB is built for local agentic jobs, using 2-bit quantization to squeeze into 24GB of memory. That makes it a fit for RTX and DGX Spark systems, letting you handle day-to-day agent tasks—web research, writing, code debugging—without the cloud. NVIDIA says Bolt 74B-A9B leads its class on PinchBench and ranks second on BrowseComp.

Then there’s Bolt E4B 4GB, which targets speedy edge deployments—think conversational agents and interactive characters on GeForce RTX, Jetson, and RTX/DGX Spark. Tuned for intelligence per gigabyte, it’s aimed at voice assistants and embedded interfaces. It rolls out October 23 and is meant for devices with tight memory budgets, making it practical for embedded or real-time use where speed and efficiency matter most.

All of these join the existing Nemotron roster: Lightning (handles high-throughput), Ultra (for complex reasoning), and the original Bolt (for edge intelligence). So now, Nemotron stretches from high-powered data centers to consumer hardware, giving organizations a menu of options—without locking them out of openness or fine-tuning.

Why This Matters: Open, Customizable AI for Real-World Business Agents

This is a big swing from NVIDIA for anyone who wants AI agents tailored to their business, but doesn’t want to be boxed in by closed, vendor-controlled models. With open weights and datasets, you can tweak and post-train these models to fit your own tools, data, and workflow—whether you’re handling insurance claims, financial analysis, healthcare, or SaaS. For example, insurance teams could run fraud-detection in-house, keeping data private.

Banks might deploy credit analysis agents that reflect their own lending rules, without sending a byte to the cloud.

The vision-language chops in Nemotron 3.5 Super VL are a real win for any agent working with real-world docs. It reads web pages, pulls out text, deciphers graphs and tables, and returns structured insights. Since it can handle both text and visuals in one go, you don’t have to bounce data between separate models—meaning faster results and better context as you move from document to decision.

Imagine an agent parsing a PDF contract, finding the important clauses, and cross-checking them with charts in the attachments—all in one inference pass.

Some organizations already customizing Super VL include:

  • Scale AI, for Skydance’s vendor credit finance agent and EY’s M&A due diligence agent
  • AfterQuery, for Hebbia’s credit agreement analysis
  • AI21 and Alice, for cybersecurity attack resilience
  • Mercor, for software coding agents using APEX-SWE benchmarks
  • Ramp, for financial document intelligence
  • Dream, for AI-powered cybersecurity research
  • Lila Sciences, for scientific discovery
  • R1, for medical claims analysis
  • TrainLoop, for dermatology data applications, including Nolla’s AI-driven prescription writing

Major software platforms—Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce, and Synopsys—are also folding Super VL into tasks like agentic search and chip design verification. Atlassian, for instance, is using it to power search across Jira and Confluence, while Salesforce is automating intricate government workflows in secure settings via Missionforce Operations. You can see how these open models adapt to industry-specific needs in real production, not just demos.

How to Use These Models in Your Business

If you’re eyeing Nemotron 3.5 models for your own AI agents, here’s a straightforward path:

  1. Grab the open weights for Super VL or Bolt from Hugging Face, ModelScope, or Open Router. You’ll find the official docs at NVIDIA’s Nemotron page. These repositories will also point you to community help and how-tos.
  2. Tune the model using NVIDIA NeMo, Python pipelines, or open datasets. NVIDIA is dropping datasets on Hugging Face that cover tool use, finance, instruction following, long context, fairness, and vision-language. You can fine-tune with your private data and tools to get the best results for your field.
  3. Deploy on whatever hardware suits your scale—DGX Spark, DGX Station, RTX PRO, GeForce RTX PCs, or Jetson. Bolt models target edge and workstations; Super VL is best for high-volume, enterprise jobs. For developers, Bolt 74B-A9B 24GB plays nice with agent harnesses and popular tools like llama.cpp, Ollama, LM Studio, Unsloth, Canonical, and vLLM, so you’re not boxed into a corner.
  4. In production, consider NVIDIA NIM, which bundles the model as a microservice with tuned configs and regular performance/security updates. NIM brings enterprise-level deployment—security patches every two weeks, performance tweaks, and capacity to serve up to 8x more users with 85% lower latency on NVIDIA Dynamo backends.
  5. Tap into the community: technical blogs, architecture guides, video tutorials, livestreams, and forums. NVIDIA’s resources can help you move faster and avoid common pitfalls as you roll out these models.

How Nemotron 3.5 Super VL Was Built and Trained

Nemotron 3.5 Super VL’s training involved roughly 26 trillion tokens—covering both initial and continued pretraining—then post-training with about 3.8 million samples, plus reinforcement learning across dozens of environments. This hefty regimen aims to build strong multimodal reasoning, instruction following, and tool use. NVIDIA is also releasing open datasets on Hugging Face for everything from tool use and finance to vision-language and fairness, as well as RL environments in NeMo Gym.

That means you can replicate or extend the training for your own custom use cases.

If you’re developing or researching, open weights, datasets, and training recipes mean you can control the full model lifecycle—test, experiment, and move to production, in-house or with partners, no vendor lock-in. That’s a big deal for industries with compliance or data residency rules, like healthcare, finance, and government.

My Perspective: The Upside and the Tradeoffs

Here’s my take: NVIDIA’s open approach is refreshing. Open weights, data, and recipes let you actually own your AI deployment—not something you get with most cloud models. You can post-train on your data, keep things on-prem, and tick the boxes for security and compliance—especially important for regulated sectors or anyone with sensitive workflows.

But you should know the tradeoffs. Running the bigger models locally means you need serious hardware—DGX Spark, RTX PRO, or similar. Bolt models make edge use more accessible, but you’ll still need NVIDIA GPUs.

There’s also a learning curve: fine-tuning, deployment, and managing inference pipelines all require technical skill. And while NVIDIA touts top scores on benchmarks, your real-world results will hinge on your own data and needs.

And remember: open models are only as secure as your deployment. Strong MLOps and security practices are a must, or the perks of on-prem control could backfire. Invest in monitoring, tight access controls, and regular updates to keep things secure and reliable.

The Bottom Line

NVIDIA’s expanded Nemotron 3.5 suite brings open, customizable AI agent models within reach—if you’ve got the right gear and know-how. If controlling your data, tuning models to your business, and deploying flexibly between cloud and edge matter to you, it’s worth a look. With open access, solid performance, and support for industry-specific tuning, Nemotron 3.5 looks well-positioned for the next wave of business AI agents.