Self-evolving AI agent ‘Hermes’ runs always on NVIDIA DGX Spark

NVIDIA announced that it will officially support Nous Research's next-generation local AI agent 'Hermes' on RTX PC and DGX Spark systems. With this support, users will be able to build a powerful local AI environment that operates 24 hours a day without being dependent on a specific model or vendor.

NVIDIA supports Nous Research's AI agent 'Hermes' in a local environment on its RTX PC and DGX Spark hardware.
Hermes, which was recently released, exceeded 140,000 GitHub stars within three months of launch and was recorded as the most used agent in the world according to the global platform OpenRouter. Developed by Noose Research, this agent is not dependent on a specific vendor or model and is designed to operate 24 hours a day in a local environment. Additionally, messaging app integration and local file access are possible.
Hermes' main competitiveness can be broadly summarized into four points:
First, it has ‘self-evolving technology’ that performs complex tasks and creates and improves new technologies on its own. Second, it operates an 'independent sub-agent' that only performs specific sub-tasks, minimizing task confusion and enabling execution in smaller context windows.
Third, after stress testing all tools and plug-ins, it operates stably without debugging even in a 30 billion parameter local model environment. Fourth, it operates as an active orchestration layer rather than simply a thin wrapper, providing consistently superior performance compared to other frameworks based on the same model.
NVIDIA DGX Spark is a system designed to sustain this agentic workflow all day. It provides 128GB of integrated memory and 1 petaflop AI performance, enabling continuous execution of expert mixture (MoE) models with 120 billion parameters. Users can easily link and build a local agent in the Hermes GitHub repository using llama.cpp, LM Studio, Ollama, etc.
Meanwhile, NVIDIA also announced additional updates to expand the RTX AI PC ecosystem. It provides the NVFP4 checkpoint of Google's Gemma 4 model, supporting up to 3 times faster inference speed on Blackwell GPU, and updated compatibility with Mistral Medium version 3.5. In addition, the company released a new open source stack, 'NemoClaw', which optimizes the OpenClaw environment by supporting the Windows environment (WSL2).