Build Muse Glimmer 30B with Hermes and ExecuTorch on DGX Spark
Introduction
Serve Muse Glimmer locally on the GPU
Start Hermes on the CPU with restricted local tools
Use Hermes with Muse Glimmer as a coding agent
Extend your local Hermes and Muse Glimmer agent
Next Steps
Build Muse Glimmer 30B with Hermes and ExecuTorch on DGX Spark
Who is this for?
Developers who want to run a local AI agent on NVIDIA DGX Spark using the Hermes Agent framework and the Muse Glimmer 30B parameter model.
What will you learn?
After completing this Learning Path, you'll be able to perform the following tasks:
- Serve Muse Glimmer 30B with text and image input through a local ExecuTorch CUDA endpoint.
- Connect Hermes Agent on the Arm-based CPU to the local model server.
- Restrict agent tools to a dedicated workspace and a network-disabled container.
- Use image and text inputs to diagnose a release error and verify the agent-generated code repair against acceptance tests.
Prerequisites
Before starting, you'll need the following prerequisites:
- SSH access to an NVIDIA DGX Spark running Ubuntu 24.04-based DGX OS, NVIDIA driver 580 or later, and the CUDA 13.0 development toolkit
- Docker Engine installed and permission to run containers
- Python 3.12 with venv and pip support, curl, Git, as well as C++17 compiler, Make, and CMake 3.26 or later
- At least 100 GiB of available disk space and 80 GiB of available system memory before the native model export
- Internet access to download services during setup
Summary
This summary was drafted with an approved AI-assisted workflow and reviewed by Arm contributors before publication. Human technical review remains part of the process so the final page reflects engineering rigor, accuracy, and Arm editorial standards.
You’ll serve Muse Glimmer 30B through a local ExecuTorch CUDA endpoint, then connect Hermes on the Arm-based CPU. First, you’ll verify model access and the restricted Docker sandbox. Then, you’ll provide a release-summary image so that Hermes can identify the incorrect deployment decision, inspect the Python checker, and repair it. You’ll run all four acceptance tests and review the code diff. Finally, you’ll identify ways to extend the agent and add more capabilities.