Hermes, a Self-Improving AI Agent Built on NVIDIA RTX PCs and DGX Spark

Hermes Agent is an open-source agent framework made by Nous Research that connects messaging apps with local files and applications, running around the clock. Its GitHub stars passed 140,000 in under 3 months, and according to OpenRouter, it was introduced as the most-used agent in the world last week.

It uses a vendor- and model-agnostic structure, making it well suited to local hardware like NVIDIA RTX PC, RTX PRO workstations, and DGX Spark. Its key differentiators are the following four points.

• Self-evolving skills: Every time it handles a complex task or receives feedback, it saves what it learned as a skill, improving itself over time.
• Contained sub-agents: It splits subtasks into short-lived, isolated workers, allowing it to operate cleanly even with small context windows.
• Reliability built in from the design stage: Nous Research curated and stress-tested the skills, tools, and plugins, improving stability even on local models.
• Same model, better results: Even using the same model, it emphasizes acting as an orchestration layer rather than a simple wrapper, yielding better results.

In line with this, Alibaba's Qwen 3.6 open-weight LLM was presented as a model for local agents. Qwen 3.6 35B was described as running on about 20GB of memory while surpassing 120B-class models that require 70GB+, and Qwen 3.6 27B aims for accuracy comparable to 400B-class models like Qwen 3.5 397B while being 1/16 the size. It was added that NVIDIA Tensor Cores reduce inference latency, speeding up multi-step tasks and skill rewriting.

DGX Spark is positioned as a compact, standalone machine for always-on agent workflows. With 128GB of unified memory and 1 petaflop of AI performance, it can run 120B-class MoE models all day long, and the lighter Qwen 3.6 35B runs even faster, leaving headroom for concurrent tasks.

To get started with Hermes, you download it from the GitHub repository and connect it to the local model and runtime of your choice. It supports llama.cpp, LM Studio, and Ollama, and Hermes Agent provides built-in integration with LM Studio and Ollama. Additionally, it introduced up to a 3x improvement in Qwen 3.6 token generation speed on RTX PRO GPUs, Gemma 4 NVFP4 checkpoints, Mistral Medium 3.5 compatibility, and WSL2 support for NemoClaw.