NVIDIA’s latest Nemotron releases address two deployment settings: specialized enterprise workflows and agents running on personal computers.
Released on October 15, Nemotron 3.5 Super VL combines visual understanding with reasoning and tool use. Its mixture-of-experts architecture contains 120 billion parameters, activating 12 billion per token.
The model can work across text, images, documents and video. NVIDIA also provides customization resources for teams adapting it to their own terminology, tools and operating procedures.
A document-processing agent, for example, may need to interpret a table, compare information across sources and follow an organization’s rules. That illustrates the intended workflow; successful deployment still requires testing against the actual documents and decisions involved.
For local use, NVIDIA released Nemotron 3.5 Bolt 74B-A9B 24GB. Its model card describes a vision-language model using native two-bit quantization to fit within a 24 GB memory footprint.
The distinction matters for developers choosing between the releases. Super VL emphasizes adaptable multimodal workflows, while this Bolt version targets constrained local environments. Application memory, context length and other running software remain part of deployment planning.
NVIDIA reports efficiency and accuracy gains on selected workloads. Those results require attention to the hardware, customization and evaluation conditions behind each comparison.
Sources: https://developer.nvidia.com/blog/