Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models
NVIDIA unveiled three new Nemotron 3.5 models, including the open 120B Super VL and edge-focused Bolt variants, to optimize agentic workloads from data centers to local devices.
NVIDIA introduced Nemotron 3.5 Super VL, an open 120-billion-parameter model designed for specialized agentic tasks, offering the fastest task completion and highest accuracy in its class. The model excels in instruction following, tool use, and visual understanding, performing comparably to larger models while using roughly one-fifth the parameters. It is optimized for enterprise deployments and supports post-training on proprietary data, enabling organizations to customize agents for workflows such as fraud detection or credit analysis without cloud dependency.
Two new edge models, Nemotron 3.5 Bolt 74B-A9B 24GB and Bolt E4B 4GB, were also released to balance intelligence and speed for local agentic workloads. Bolt 74B-A9B 24GB runs on NVIDIA RTX/DGX Spark with 2-bit quantization, fitting within 24GB of memory while handling tasks like web research and code debugging locally. Bolt E4B 4GB, launching October 23, targets physical and conversational agents on GeForce RTX, Jetson, and RTX/DGX Spark, enabling fast local voice assistants and embedded interfaces.
The Nemotron family now includes Super VL, Bolt, Lightning, and Ultra, each tailored for specific workloads. Super VL is being post-trained by organizations such as Scale AI, EY, and AI21 for specialized tasks like M&A due diligence and cybersecurity resilience. The model’s multimodal capabilities allow agents to process visual documents, web pages, and PDFs without routing tasks between separate systems, improving efficiency in workflows like medical claims analysis and scientific discovery.
NVIDIA released open weights, datasets, and training recipes for Nemotron 3.5 Super VL, alongside reinforcement learning environments in NeMo Gym. The model was trained on 26 trillion tokens and post-trained with 3.8 million samples, while open datasets on Hugging Face cover agentic tool use, finance, and vision-language understanding. Organizations can deploy the model on-premises to meet security and data-residency requirements, ensuring data privacy while adapting the model to proprietary workflows.