Nemotron 3 Ultra on DGX Station: agent setup explained

This session explains a local AI workflow built around Nemotron 3 Ultra on a GB300 DGX Station running Ubuntu. The presenters use vLLM to serve the model and an RTX Pro 6000 for Blender rendering. They describe Ultra as having 550 billion total parameters and 55 billion active parameters, and report close to 230 GB of memory use for the demonstrated vLLM engine. CPU offloading helps accommodate a model that does not fit entirely in high-bandwidth memory.

Hermes is the AI agent inside an OpenShell sandbox. Its skills connect it to Blender through MCP, where it can inspect a scene and request changes. The example replaces cubes with rubber spheres and uses Omniverse PhysX integration to simulate them bouncing down stairs. The presenters show a prepared final simulation alongside the live scene inspection. OpenShell policies specify filesystem access and permitted network endpoints; Blender actions depend on what the MCP service exposes.

The demonstration also includes Codex as a coach for harder tasks, so the combined workflow uses an external service alongside the local model. The speaker says Hermes saves reusable skills and needs less coaching on similar follow-up requests.

For developers who want to run models locally, the performance advice is specific: use the recipe and launch flags validated for the hardware. The discussion covers the tradeoff between single-user speed and concurrent throughput, plus NVIDIA Dynamo for request routing and prefix-cache management.