The additions expand Nvidia's open-model portfolio across enterprise infrastructure, desktop computers and embedded systems. They join the existing Nemotron Lightning model, which focuses on processing large volumes of requests, and Nemotron Ultra, which is designed for demanding reasoning tasks.
Nvidia is positioning Super VL as its primary option for organizations developing specialized AI agents that must interpret information, follow instructions and execute multistep workflows. According to the company, the model delivers performance comparable to its largest reasoning model on several multimodal tasks while requiring approximately one-fifth as many parameters.
Nvidia also claims Super VL can process certain agent workloads up to five times faster than competing open models in its category. The performance figures are based on benchmarks cited by the company, including results from the Artificial Analysis Intelligence Index.
The model's capabilities extend beyond text-based reasoning. Super VL can examine images, interpret charts and tables, extract information from PDFs and analyze web pages. Its combined language and visual processing capabilities allow agents to work with different types of information without transferring tasks between separate models.
That functionality is intended for enterprise workflows involving documents and structured information. For example, an AI agent could review financial records, identify relevant figures and apply an organization's internal criteria without requiring a separate system to interpret visual content.
Nvidia says dozens of organizations are already adapting Super VL for specialized applications using their own datasets, policies and business processes.
Scale AI has customized the model for a vendor credit finance agent used by Skydance and a mergers and acquisitions due diligence agent for EY. AfterQuery has adapted it for Hebbia to analyze credit agreements, while Ramp is applying the technology to financial document processing.
Other organizations are targeting software development, cybersecurity, healthcare and scientific research. Mercor has trained the model using the APEX-SWE coding benchmark to support agents that plan, generate and troubleshoot software. AI21 and Alice are working on cybersecurity applications, while Lila Sciences is using the model for scientific research and information retrieval.
R1 is customizing Super VL for medical claims processing and revenue cycle management. TrainLoop has also trained the model using dermatology data from Nolla, which Nvidia identifies as the first company approved to use AI for writing initial patient prescriptions.
Several enterprise software providers are incorporating Super VL into their products or development efforts. Atlassian is using the model for AI-powered searches across applications such as Jira and Confluence, while CrowdStrike is adapting it for automated defensive cybersecurity operations.
Salesforce is customizing Nemotron for Missionforce Operations, targeting government workflows that require secure environments. Dassault Systèmes is applying the technology to engineering design and simulation, and Perplexity is integrating it into its search and agent products. Synopsys has also established interoperability with the model for semiconductor design verification.
Alongside Super VL, Nvidia introduced the Nemotron 3.5 Bolt series to support AI agents operating on devices with more limited computing resources.
The larger Bolt 74B-A9B 24GB model uses native 2-bit quantization to operate within 24GB of memory. Nvidia designed it for systems including GeForce RTX computers and DGX Spark, allowing users to run agent workloads locally while continuing to use other applications.
The model supports tasks such as researching information online, writing software and debugging code without transmitting the underlying workloads to cloud infrastructure.
Nvidia reports that Bolt 74B-A9B ranks first among comparable open models on PinchBench and second on BrowseComp. The company describes its performance as a compromise between the processing speed of smaller models and the reasoning capabilities of more computationally demanding alternatives.
The smaller Bolt E4B 4GB targets conversational AI and physical agents operating on hardware such as GeForce RTX and Jetson systems. It is designed to handle extended interactions and agent-driven tasks for applications including voice assistants, interactive digital characters and embedded interfaces.
For organizations concerned about data security and deployment flexibility, Nvidia is emphasizing the ability to customize and operate these models within their own infrastructure.
Nemotron 3.5 Super VL is distributed with publicly available model weights, training datasets and instructions for reproducing its training processes. Developers can modify the model using Nvidia NeMo, Python-based tools and additional open datasets.
The company says Super VL underwent initial and continued training on approximately 26 trillion tokens, followed by additional training involving roughly 3.8 million samples and reinforcement learning across multiple environments.
Nvidia is also publishing datasets covering tool use, financial applications, instruction following, long-context processing, fairness and visual understanding. Reinforcement learning environments are available through NeMo Gym.
For production deployments, Nvidia offers Super VL through its NIM microservice platform, which provides optimized configurations and enterprise support. The company says NIM receives security and performance updates every two weeks.
Nvidia also claims that deploying Super VL with its Dynamo inference software can increase the number of users served by as much as eight times while reducing latency by 85%.
Super VL is available through Hugging Face, OpenRouter, the ModelScope community and Nvidia NIM. Supported hardware includes DGX Spark, DGX Station, RTX and RTX PRO systems, Jetson devices and data center infrastructure.
Bolt 74B-A9B is available on Hugging Face with support for development and deployment tools including llama.cpp, Ollama, LM Studio, Unsloth and vLLM. It is also designed to run on Nvidia RTX Spark laptops arriving October 16, as well as compatible desktop and workstation systems.
The releases give developers several options within the Nemotron family depending on the complexity of their workloads and the hardware available. Super VL targets enterprise agents that require advanced reasoning and visual understanding, while the Bolt models prioritize local execution on devices with tighter memory constraints.
By offering the models with open weights and customization tools, Nvidia is giving organizations the option to adapt AI agents to specific workflows while maintaining control over where their models operate and how their data is processed.
This analysis is based on reporting from NVIDIA.
Image courtesy of NVIDIA.
This article was generated with AI assistance and reviewed for accuracy and quality.
