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

NVIDIA has widened its Nemotron family with three open models that split cleanly by where the work happens. Nemotron 3.5 Super VL is the big one: a 120-billion-parameter model aimed at specialized always-on agents. Alongside it come Nemotron 3.5 Bolt 74B-A9B 24GB for local agentic workloads and Nemotron 3.5 Bolt E4B 4GB for fast conversational agents. Super VL and Bolt 74B-A9B are out now; Bolt E4B lands on Oct. 23.

The pitch is simple enough. Super VL is meant for heavy enterprise tasks, while the Bolt models bring that same agent idea down to local hardware and edge devices. NVIDIA is selling this as coverage across the full stack, from data center systems to desktop and embedded gear, so companies do not have to pick between capability and deployability. Super VL joins Lightning and Ultra in the family, with Lightning for high-throughput work and Ultra for harder reasoning.

Super VL is the headliner because NVIDIA is pushing its efficiency story hard. Built with the Nemotron Coalition, it is supposed to match NVIDIA’s largest reasoning model on multimodal agentic tasks such as instruction following, tool use, multistep business workflows and visual understanding, but at roughly one-fifth the parameter count. NVIDIA also says it is the fastest open agentic model in its class, and that on the Artificial Analysis Intelligence Index it can reach up to 5x the throughput of other models in that class. The point is not just speed for speed’s sake. It is speed for workflows that stretch across days or weeks.

The company is also trying to make the openness angle feel practical, not ideological. Super VL ships with open weights, open datasets and training recipes, and NVIDIA says organizations can post-train it on their own data and keep it on their own infrastructure. It is already being customized across a long list of use cases, from Skydance’s vendor credit finance agent to EY’s M&A due diligence agent, plus cybersecurity, coding, financial documents, medical claims and more. NVIDIA says 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.

The Bolt models fill in the local side. Bolt 74B-A9B 24GB uses native 2-bit quantization to fit in 24GB of memory and is designed for NVIDIA RTX/DGX Spark, while Bolt E4B 4GB targets physical and conversational agents on GeForce RTX, Jetson and RTX/DGX Spark. NVIDIA is clearly betting that agent software should not have to bounce to the cloud for every research query, draft or debugging job. That is the real message here: the company wants open models to feel like infrastructure, not a demo.