Hermes: the AI Agent that evolves on its own, powered by NVIDIA RTX and Qwen 3.6
AI Agents are changing how people get work done every day, and the open source community is accelerating that shift with developments worth paying attention to. After the success of OpenClaw, the community embraced new agentic open source frameworks, and the latest example is Hermes Agent, developed by Nous Research.
The project crossed the 140,000 stars mark on GitHub in less than three months and, according to OpenRouter data, is already considered the most widely used agent in the world. Numbers like that don’t just happen out of nowhere.
Hermes was built on two pillars that have historically been hard to achieve in AI agents: reliability and self-improvement — meaning the ability to learn and evolve on its own over time. It is provider and model agnostic by design, which means it works with different LLMs and services without any lock-in.
And when you combine this framework with Qwen 3.6 models from Alibaba, running on NVIDIA RTX, RTX PRO, or DGX Spark hardware, the result is an extremely powerful local stack — no cloud dependency, no unnecessary latency, and data center-level performance in your own environment.
That’s exactly the combination we’re going to break down here. 🚀
What makes Hermes different from other AI Agents
Most AI Agent frameworks available today handle a specific problem well enough but stumble when the context changes, when there are intermediate failures, or when the agent needs to make decisions that weren’t part of the original flow. Hermes was designed specifically to overcome these limitations. Like other popular agents, it integrates with messaging apps, accesses local files and applications, and runs around the clock. But four standout capabilities set it apart from everything else out there right now.
Self-Evolving Skills: abilities that improve on their own
Hermes writes and refines its own skills. Every time the agent encounters a complex task or receives feedback, it saves what it learned as a skill, allowing it to adapt and improve over time. This continuous learning cycle means that the more you use Hermes, the more precise, efficient, and aligned with your goals it becomes — without requiring manual retraining or constant intervention.
Contained Sub-Agents: isolated, focused workers
Hermes treats sub-agents as isolated, short-lived workers, each dedicated to a specific subtask with focused context and a defined set of tools. This keeps task organization clean, minimizes confusion for the agent, and allows Hermes to work with smaller context windows — which is ideal for models running locally.
Reliability by design
Nous Research curated and thoroughly tested every skill, tool, and plug-in that ships with Hermes. The result is a framework that simply works, even with local models in the 30 billion parameter range, without the constant debugging that most other agent frameworks demand. For anyone who has spent hours trying to get an agent to run properly, this difference is huge.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Same model, better results
Comparisons by developers using identical models across different frameworks show consistently better results with Hermes. The difference lies in the framework itself: Hermes acts as an active orchestration layer, not a thin wrapper on top of the model. This enables persistent, on-device agents instead of task-by-task executions.
Another factor behind the project’s explosive growth is the way Nous Research led development in an open and collaborative manner. The community actively contributed integrations, fixes, and new features, which accelerated the framework’s maturity cycle in impressive fashion. Reaching 140,000 stars on GitHub in less than three months isn’t just a flashy number for a slide deck — it’s a real barometer of adoption and trust from the technical community, which tends to be quite critical of tools that promise more than they deliver.
Qwen 3.6: data center intelligence running locally
When Alibaba released Qwen 3.6, the goal was clear: deliver high-performance, efficient language models with advanced reasoning capabilities that could also run viably on local hardware. This new series of open weight LLMs continues the acclaimed Qwen 3.5 lineup, representing another leap forward for local AI Agents.
The numbers are impressive. The new Qwen 3.6 35B model runs on roughly 20GB of memory and still outperforms 120 billion parameter models that require over 70GB of memory. Meanwhile, the Qwen 3.6 27B is a dense model with more active parameters that matches the accuracy of massive models like Qwen 3.5 397B while being one-sixteenth the size. On high-performance RTX GPUs, this model has the computational horsepower needed to deliver a fast and responsive experience.
From a technical standpoint, Qwen 3.6 features an architecture that balances parameter count with fast inference capability nicely, which is critical when you’re running locally and want real-time responses without sacrificing quality. It also supports long context windows, which allows Hermes to maintain a richer reasoning history between execution steps. This makes a massive difference in tasks involving multiple steps, dependencies between actions, and the need to revisit earlier decisions during a workflow.
The integration between Qwen 3.6 and Hermes is no accident. The way the model processes structured instructions, tool formats, and typed outputs fits perfectly with how the framework expects to receive and interpret responses. This significantly reduces parsing error rates, which is one of the main failure points in AI Agent pipelines when the base model wasn’t optimized for this type of interaction. This structural compatibility is what makes the stack work so well in practice — not just on paper. 🎯
NVIDIA Tensor Cores accelerate AI inference to deliver higher throughput and lower latency, allowing Hermes to work through multi-step tasks or refine one of its own skills in seconds, not minutes.
DGX Spark: the always-on computer for AI Agents
Agents like Hermes are built to run continuously — responding to requests, planning multi-step tasks, executing autonomously, and self-improving. The NVIDIA DGX Spark is the ideal companion for this scenario: a compact, efficient, standalone machine built for sustained agentic workflows throughout the entire day.
With 128GB of unified memory and 1 petaflop of AI performance, DGX Spark can run 120 billion parameter mixture-of-experts models all day long without breaking a sweat. And with the new Qwen 3.6 35B delivering equivalent intelligence in a leaner package, models run faster and free up capacity for users to handle concurrent workloads.
For those looking to maximize performance and ease of use, NVIDIA has released a dedicated Hermes playbook for DGX Spark, along with hands-on sessions in the agentic AI Build It Yourself series, where you can learn to build autonomous AI Agents with NemoClaw and OpenShell.
DGX Spark is already available for purchase through NVIDIA’s manufacturing partners, offering a real option for anyone who wants a true personal AI supercomputer on their desk.
NVIDIA RTX and local execution: the stack that closes the loop
Running powerful language models locally is no longer a distant dream for anyone with access to a latest-generation NVIDIA RTX. The RTX family offers native support for accelerated inference via CUDA and TensorRT, and both Hermes and the base model are built to run locally, meaning hardware quality directly determines the quality of the user experience.
Getting Hermes up and running on NVIDIA hardware is straightforward. The GitHub repository has everything you need, and you just pair it with your preferred runtime. Hermes can run alongside Qwen 3.6 via llama.cpp, LM Studio, or Ollama. The framework comes with built-in support for LM Studio and Ollama, providing the simplest path to having a local agent running in minutes.
NVIDIA RTX also benefits from specific optimizations for model quantization, such as the GGUF format combined with backends like llama.cpp, which allow you to run compact versions of Qwen 3.6 with minimal quality loss and significant speed gains. On GPUs like the RTX 4090 or RTX 3090, you can run the model at speeds that make the experience smooth enough for continuous use in production tasks — not just for experimentation. This completely changes the cost-benefit equation for developers and companies that previously relied exclusively on pay-per-token APIs.
When you close that triangle — Hermes as the agent framework, Qwen 3.6 as the base model, and NVIDIA RTX as the inference infrastructure — you have a stack that competes head-to-head with cloud-based solutions, but with full control over your environment, no unpredictable variable costs, and the flexibility to customize every layer of the architecture. 💡
Self-improvement in practice: what it actually means for users
The concept of self-improvement in AI Agents might sound abstract at first, but in practice it translates into something very concrete: an agent that makes fewer mistakes over time and learns to execute specific tasks more efficiently without you having to manually reprogram anything. In the context of Hermes, this happens through internal reflection mechanisms, where the agent analyzes the output of each action, compares it against the expected objective, and adjusts its strategy for future executions. It’s like having an assistant that not only does what you ask but also learns from its mistakes without needing you to explain everything all over again.
This behavior has important implications for designing workflows with AI Agents. Instead of building rigid pipelines with manual fallbacks for every possible scenario, you can create more flexible flows and trust that the agent will adapt when conditions change. This reduces maintenance time, decreases the number of edge cases that need to be anticipated in the initial design, and increases the overall resilience of the system. For development teams, this translates to real time savings and lower operating costs for AI-based automations over time.
What makes Hermes’ self-improvement especially interesting is that it doesn’t depend on retraining the base model to happen. The evolution occurs at the agent level, through episodic memory, prompt tuning, and dynamic strategy selection. Qwen 3.6 remains the same model, but Hermes learns to use it in increasingly smarter ways. This separation between the model and the agent logic is a sophisticated architectural decision and part of the reason Hermes pairs so well with different base models, adapting to what each one does best. 🧠
What else is happening in the RTX AI Garage ecosystem
Beyond the Hermes + Qwen 3.6 combo, NVIDIA’s local AI ecosystem continues to receive noteworthy updates worth mentioning.
NVIDIA RTX PRO GPUs are delivering up to 3x faster token generation when running Qwen 3.6 models with llama.cpp, providing the real-time responsiveness that agents need to work through multi-step tasks and maintain continuous workflows.
Google’s Gemma 4 26B and 31B models are now available as NVFP4 checkpoints for even faster performance on NVIDIA Blackwell GPUs. Combining these checkpoints with Google’s new Multi-Token Prediction drafters, you can get up to 3x faster inference speeds with identical output quality, enabling frontier-level reasoning running locally.
Mistral Medium version 3.5 also received compatibility updates for llama.cpp and Ollama, allowing users to run it on RTX PRO and DGX Spark systems.
NVIDIA also recently introduced NemoClaw, an open source stack that optimizes OpenClaw experiences on NVIDIA devices, boosting security and supporting local models. NemoClaw now supports Windows Subsystem for Linux (WSL2), bringing compatibility to enthusiasts and developers on the Microsoft platform.
The combination of Hermes, Qwen 3.6, and NVIDIA RTX represents one of the most exciting moments in the open source AI Agent ecosystem, where performance, autonomy, and local control finally come together in an accessible, well-documented stack.
What’s happening here goes well beyond a new tool or a model with impressive benchmarks. It’s the consolidation of an approach where truly autonomous AI Agents with self-improvement capabilities can run locally on accessible hardware, using high-quality open source models like Qwen 3.6, in a reliable framework like Hermes, accelerated by the inference infrastructure of NVIDIA RTX and DGX Spark. For anyone closely following the evolution of artificial intelligence, this is the kind of convergence that happens rarely and tends to redefine what’s considered possible.
