Nemotron 3 Ultra setup with NemoHermes and OpenCode

This tutorial walks through two agent setups for NVIDIA Nemotron 3 Ultra: Hermes through NemoHermes, and OpenCode directly. The supplied metadata describes a sparse mixture-of-experts model with 550 billion total parameters and 55 billion active parameters. The speaker describes its weights, data and training recipe as open.

The Hermes setup uses an installation script and an environment variable to select Hermes. Onboarding then selects NVIDIA endpoints and accepts an NVIDIA API key. After choosing Ultra for inference, the presenter connects to a sandbox and launches the AI agent inside the OpenShell runtime. This demonstration uses build.nvidia.com for inference; it does not show how to run models locally or establish offline operation.

For the research example, Hermes retrieves Karpathy's NanoChat, sets it up and performs smoke tests. The task stops after one optimization rather than continuing in a persistent loop. The presenter reports a small improvement, but the transcript gives no numerical result or independent validation.

The OpenCode section shows configuration through a JSON block and mentions configuration files as another option. The coding assistant receives a prompt to create a Newton's cradle in JavaScript. Ultra then starts a server so the presenter can open the resulting demo through a link.