Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models

NVIDIA has significantly broadened its AI portfolio today with the launch of the Nemotron 3.5 Super VL, a powerful 120-billion-parameter open model engineered to deliver industry-leading speed and precision for specialized, always-on agentic workflows. Alongside this flagship release, the company introduced two new edge-focused models: the Nemotron 3.5 Bolt 74B-A9B 24GB, designed to optimize the balance between intelligence and speed for local workloads, and the Nemotron 3.5 Bolt E4B 4GB, which focuses on high-efficiency conversational capabilities.

These additions to the Nemotron family—which already includes the high-throughput Lightning and the reasoning-focused Ultra—ensure that organizations now have a comprehensive suite of open models capable of scaling from massive enterprise data centers down to local edge hardware. By providing these tools, NVIDIA is removing the traditional barrier between high-level capability and practical, on-premises deployability.

Nemotron 3.5 Super VL: Frontier-Level Intelligence at Scale

The Nemotron 3.5 Super VL represents a major leap forward in multimodal performance. Developed in partnership with the Nemotron Coalition, the model excels at complex agentic tasks, including multi-step business workflows, tool utilization, nuanced instruction following, and sophisticated visual analysis.

Remarkably, the model achieves this frontier-level accuracy while utilizing only about one-fifth of the parameter count of NVIDIA’s largest reasoning models. This efficiency makes it the fastest open model in its category, boasting throughput speeds up to five times higher than its closest competitors, according to the Artificial Analysis Intelligence Index leaderboard.

"Unlike closed frontier models, organizations can post-train Super VL on their own data, deploy it on their own infrastructure and oversee what they build."

By keeping the model open, NVIDIA empowers enterprises to post-train the system on proprietary data and deploy it within their own secure environments, ensuring that sensitive information remains private and under the organization's direct control.

Bringing Agentic Intelligence to the Edge with Bolt

While Super VL targets high-volume enterprise tasks, the new Bolt models are designed to bring sophisticated AI capabilities to local hardware, such as NVIDIA RTX-powered systems, Jetson modules, and DGX Spark stations.

• Nemotron 3.5 Bolt 74B-A9B 24GB: This model utilizes native 2-bit quantization to fit within 24GB of memory. It is engineered to handle everyday agentic tasks—such as web research, coding assistance, and debugging—locally, eliminating the need to send sensitive data to the cloud. It currently ranks as the #1 open model in its class on PinchBench and #2 on BrowseComp.
• Nemotron 3.5 Bolt E4B 4GB: Launching on October 23, this model is optimized for physical and conversational agents. It is specifically built to power fast, responsive voice assistants and interactive interfaces on hardware like GeForce RTX PCs and Jetson devices.

Industry-Wide Adoption and Customization

The versatility of the Nemotron 3.5 Super VL has already sparked widespread adoption across diverse sectors. Organizations are leveraging the model’s open nature to fine-tune it for highly specialized tasks.

Key industry applications include:

• Finance & Legal: Scale AI is utilizing the model for vendor credit finance agents, while EY is applying it to M&A due diligence. Ramp is using it for financial document intelligence, and R1 is deploying it for medical claims analysis.
• Cybersecurity: AI21 and Alice are using the model to bolster attack resilience, while CrowdStrike is customizing it for agentic blue teaming.
• Software & Engineering: Mercor is using it for coding agents that plan and debug software, and Synopsys is integrating it for semiconductor design verification.
• Scientific Discovery: Lila Sciences is applying the model to accelerate research and data search.
• Healthcare: TrainLoop has post-trained the model with dermatology data from Nolla, which recently achieved a milestone as the first company approved for AI-generated initial patient prescriptions.

Furthermore, major software platforms like Atlassian, Perplexity, Salesforce, and Dassault Systèmes are integrating the model to enhance agentic search, engineering design, and complex workflow automation.

Vision-Language Strengths: Seeing the Full Picture

A standout feature of the Nemotron 3.5 Super VL is its advanced vision-language processing. It is specifically trained to interpret the types of visual data agents encounter in real-world business environments, such as:

• Optical Character Recognition (OCR): Accurately extracting text from complex documents.
• Structured Data Analysis: Interpreting tables, charts, and graphs within PDFs and web pages.
• Contextual Continuity: By combining text and vision in a single model, the agent maintains context as it moves from document analysis to data synthesis and final decision-making, eliminating the need for inefficient task routing between disparate systems.

Security, Privacy, and Performance

NVIDIA’s commitment to openness extends beyond the model weights. The company is providing open datasets and training recipes, allowing organizations to maintain complete sovereignty over their AI infrastructure.

For enterprise-grade requirements, NVIDIA NIM packages the model as a production-ready microservice. This includes validated configurations and regular updates, ensuring that security patches and performance optimizations are delivered every two weeks. When paired with NVIDIA Dynamo, organizations can achieve up to eight times higher user capacity with an 85% reduction in latency.

Technical Foundation

The development of Nemotron 3.5 Super VL involved a massive training regimen:

• Pretraining: Approximately 26 trillion tokens.
• Post-training: Roughly 3.8 million samples.
• Reinforcement Learning: Conducted across dozens of specialized environments using NeMo Gym.

Availability and Deployment

The Nemotron 3.5 Super VL is available immediately via Hugging Face, Open Router, the ModelScope community, and as an NVIDIA NIM microservice. It is designed to run across the entire NVIDIA hardware stack, from desktop-based RTX systems to the most powerful data center configurations.

The Nemotron 3.5 Bolt 74B-A9B 24GB is also available now on Hugging Face, with broad support for popular local inference tools including llama.cpp, Ollama, LM Studio, Unsloth, and vLLM.

As the AI landscape shifts toward more specialized, agentic, and localized deployments, NVIDIA’s latest Nemotron releases provide the necessary infrastructure for developers and enterprises to build, customize, and scale their own intelligent agents with unprecedented control and efficiency.