Nvidia launches Nemotron 3.5 open models for agents

Nvidia has released Nemotron 3.5 Super VL and two new Nemotron 3.5 Bolt models, expanding its open model range for AI agents across servers, desktop systems and edge devices.

Super VL is a 120-billion-parameter open model aimed at specialised always-on agents. Bolt 74B-A9B 24GB and Bolt E4B 4GB target smaller local and edge deployments. Super VL and Bolt 74B-A9B are available now, while Bolt E4B will follow later.

The release extends Nvidia's Nemotron family, which also includes Lightning and Ultra. Nvidia positions Lightning for high-throughput workloads, Ultra for more demanding reasoning tasks and Bolt for edge use.

Model range

Super VL is designed for multimodal agent tasks such as instruction following, tool use, multistep business workflows and visual understanding. Nvidia says it delivers accuracy comparable with its largest reasoning model at about one-fifth of the parameter count.

Nvidia also describes Super VL as the fastest open agentic model in its class, citing throughput of up to five times that of rivals on the Artificial Analysis Intelligence Index leaderboard. The model is open for post-training on proprietary data and can be deployed on an organisation's own infrastructure.

Bolt 74B-A9B 24GB is designed for local workloads on Nvidia RTX and DGX Spark systems. It uses native 2-bit quantisation to fit into 24GB of memory, allowing it to run alongside other applications on local hardware.

The model is intended for tasks such as web research, writing and debugging code without sending workloads to cloud services. Nvidia cites rankings of first among open models in its class on PinchBench and second among open models on BrowseComp.

Bolt E4B 4GB is aimed at physical and conversational agents running on GeForce RTX, Jetson and RTX or DGX Spark devices. It is designed for local voice assistants, interactive characters and embedded interfaces.

Customisation push

Dozens of organisations are post-training Super VL with their own data, tools, policies and workflows, according to Nvidia. The company identified users across finance, cybersecurity, software development, healthcare and scientific research.

Scale AI has post-trained the model for Skydance's vendor credit finance agent and for EY's M&A due diligence agent. AfterQuery customised the model for Hebbia for credit agreement analysis, while AI21 and Alice post-trained it for cybersecurity attack resilience.

Mercor has used the model on the APEX-SWE coding benchmark for coding agents that plan, write and debug software. ModelScope supports fine-tuning through its ms-swift framework for specialised multimodal applications.

Other organisations named by Nvidia include Ramp for financial document intelligence, Dream for cybersecurity research and vulnerability analysis, Lila Sciences for scientific discovery and search, and R1 for medical claims analysis and revenue cycle management. TrainLoop post-trained the model with dermatology data from Nolla.

Software platform groups integrating Super VL include Atlassian for agentic search across products such as Jira and Confluence, CrowdStrike for enterprise security work, Dassault Systèmes for engineering design and simulation, Perplexity for search tasks, Salesforce for government workflow automation, and Synopsys for semiconductor design verification interoperability.

Vision and control

The model's visual strengths focus on common business documents and interfaces, including optical character recognition, tables, charts and other structured information, Nvidia says. It can process webpages, PDFs and images, extract text and return results based on the information it finds.

Nvidia says its open approach is intended to give users more control over data handling, security and deployment. Super VL ships with open weights, datasets and training recipes, and organisations can run it wherever their security and data residency requirements demand.

In practical terms, the model can be adapted for uses such as fraud review in insurance and credit assessment in financial services while keeping systems on premises. Customisation is supported through Nvidia NeMo, Python-based pipelines and open datasets.

Super VL was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained with roughly 3.8 million samples and reinforcement learning across dozens of environments. Nvidia is also releasing open datasets covering tool use, finance, instruction following, long context, fairness and vision-language understanding, along with reinforcement learning environments in NeMo Gym.

Nvidia says the model can run across its hardware range, including DGX Spark, DGX Station, RTX Pro and RTX systems, Jetson platforms, and data centre and cloud environments. Bolt 74B-A9B is also available with support for software including llama.cpp, Ollama, LM Studio, Unsloth and vLLM.

The three models are intended to cover deployments from large enterprise servers to local hardware and edge devices, giving organisations an open model option for each infrastructure tier.