NVIDIA expands Nemotron with open models for enterprise and local agents

NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter model for specialized, always-on AI agents, alongside Nemotron 3.5 Bolt 74B-A9B 24GB for local workloads. A smaller conversational model, Nemotron 3.5 Bolt E4B 4GB, is due Oct. 23. The additions extend the family across enterprise servers, desktops and edge devices.

Super VL combines text and visual understanding for instruction following, tool use and multistep business workflows. It can extract information from webpages, PDFs, tables and charts without routing tasks between separate models. NVIDIA says its accuracy is comparable with its largest reasoning model at roughly one-fifth the parameter count. Organizations can post-train it on proprietary data and deploy it on their own infrastructure.

Bolt 74B-A9B 24GB uses native 2-bit quantization to fit in 24GB of memory, supporting local research, writing and coding tasks without sending workloads to the cloud. Bolt E4B 4GB targets voice assistants, interactive characters and embedded interfaces on GeForce RTX, Jetson and RTX/DGX Spark hardware.

Companies customizing or integrating Super VL include Atlassian, CrowdStrike, Perplexity and Salesforce, with applications spanning search, cybersecurity and government workflows. NVIDIA is providing open weights, datasets and training recipes, with customization supported through NeMo. Super VL is available through Hugging Face, Open Router, ModelScope and NVIDIA NIM; the larger Bolt model is available on Hugging Face with support for local inference tools including Ollama and LM Studio.