NVIDIA has unveiled three open AI models designed to bring autonomous agents to more computing environments. The lineup centers on Nemotron 3.5 Super VL, a 120-billion-parameter model built for complex tasks. Two smaller variants target users who want to run AI agents directly on local hardware.
The release expands NVIDIA’s push toward customizable AI that organizations can operate on their own infrastructure. It also addresses a practical challenge for developers: running capable agents without sending every request to a cloud service.
Super VL targets complex workflows
Nemotron 3.5 Super VL handles tasks that require more than generating text. It can interpret visual information and follow instructions while using external tools to complete multistep workflows.
NVIDIA claims the model completes tasks five times faster than the nearest competitor in its class. The company also says its accuracy approaches that of NVIDIA’s largest reasoning model, despite having roughly one-fifth as many parameters.
Those capabilities could help enterprises automate work that involves lengthy research or detailed document analysis. Developers can further train the model on proprietary information to improve its performance on specialized assignments.
Meet NVIDIA Nemotron 3.5 Super VL
— NVIDIA AI (@NVIDIAAI) October 15, 2026
Our new open model combines reasoning and tool use with text, image, document, and video understanding for long-running AI agents.
120B total parameters with 12B activepic.twitter.com/BcbED05qSG
Several organizations have already begun customizing Super VL. Scale AI adapted it for Skydance’s vendor credit finance agent. It also worked on a separate version for EY’s merger-and-acquisition due diligence agent.
AfterQuery customized the model for Hebbia’s credit agreement analysis. Other projects focus on cybersecurity, where AI21 and Alice have worked on attack resilience. Mercor has also post-trained the model for coding tasks covered by the APEX-SWE benchmark.
Bolt brings agents on-device
Nemotron 3.5 Bolt targets hardware with more limited memory. Its two variants offer different approaches to running AI agents without depending on remote infrastructure.
The larger Bolt 74B-A9B uses native 2-bit quantization to fit within 24GB of memory. That footprint allows compatible systems to run the model alongside other applications.
Developers can use it for local coding assistance or delegate web research to an agent framework. NVIDIA ranks the model first among open models in its class on PinchBench. It also claims second place among open models on BrowseComp.
Bolt E4B has a much smaller 4GB memory footprint. NVIDIA designed it for conversational agents that need to respond quickly on local devices. Potential applications include voice assistants and interactive characters.
The model also targets physical AI systems running on NVIDIA Jetson hardware. Bolt E4B is scheduled to arrive on October 23. The larger Bolt model is already available.
Open models offer more control
NVIDIA is releasing Super VL with open weights and supporting training materials. Its training process covered approximately 26 trillion tokens. Post-training involved around 3.8 million samples, followed by reinforcement learning across dozens of environments.
Organizations can adapt the model through NVIDIA NeMo and Python-based pipelines. Local deployment also gives them greater control over sensitive information, depending on their infrastructure and configuration.
Super VL is available through Hugging Face, OpenRouter, and ModelScope. NVIDIA also offers it through NIM, its platform for deploying production AI services. Bolt 74B-A9B is available on Hugging Face. Developers can run it using tools such as Ollama or LM Studio.
The bigger question is how these models perform outside benchmark tests. Their appeal will ultimately depend on whether local hardware can deliver reliable results without demanding excessive memory or sacrificing response speed.