NVIDIA Launches Hermes: AI agents Autoimprovable for RTX and DGX Spark PCs

NVIDIA announced Hermes, a new class of artificial intelligence agents designed to evolve on their own and operate reliably directly on user hardware. Unlike cloud-based solutions, the system is agnostic in relation to providers and models, optimized for continuous use and without relying on remote connection. The novelty represents a significant advance in the autonomy of local agents, with the potential to transform from development workflows to business applications.

What are self-improving AI agents Hermes?

Hermes are self-improvable AI agents — that is, able to learn from their own execution and refine their behavior without human intervention. They function as intelligent wizards who can perform complex tasks, interact with APIs, manage files and make context-based decisions, all running locally on PCs equipped with boards NVIDIA RTX, workstations PRO or the new DGX Spark. The focus is to offer enterprise-level performance with full data privacy.

How Hermes improves itself: the mechanism of continuous improvement

The system uses a continuous feedback mechanism: at the end of each task, the agent records his correct answers and errors, adjusting his internal model to improve future responses. This self-improvement cycle runs locally, taking advantage of NVIDIA's dedicated AI hardware to make quick inferences without sending information to external servers. According to the company, this allows agents to become more precise over time, adapting to the specific usage patterns of each user or organization.

Why does running locally matter to Brazilian companies? Privacy and performance

Local execution eliminates risks of leaking sensitive data to cloud servers, a critical factor for sectors such as finance, health and law. In addition, it reduces data transfer costs and prevents network latency, ensuring real-time responses even with unstable connections. For Brazilian developers, Hermes opens up possibilities to create smart applications that work offline or in environments with connectivity restrictions. Recently, a Startup raises $650 million to create AI that learns and improves its own code, demonstrating the growing market interest in such technologies.

Which hardware is compatible with Hermes? RTX and DGX Spark PCs

Hermes is designed to work on any device with NVIDIA RTX cards (40 and 50 series), professional RTX PRO line workstations and DGX Spark, a compact computer focused on inference of large models. NVIDIA states that the system is optimized to make the most of the Tensor Core cores and dedicated memory of these GPUs, allowing the execution of language models with billions of parameters in a fluid way. The company also provides customization tools via RTX AI Garage, where developers can download and adapt agents to their needs.

Frequently Asked Questions about self-improvable AI agents

Can Hermes agents be used without internet connection?

Yeah. As all processing occurs locally, Hermes works completely offline after initial installation, ensuring total data privacy.

Is a PC with NVIDIA RTX card required to use Hermes?

Yeah, Hermes requires core hardware. Tensor Dedicated colour available on RTX, RTX PRO and DGX lines Spark, being incompatible with conventional GPUs.

Does Hermes work with models from any provider?

Yes, the system is agnostic in relation to providers and models, and can run from open-source LLMs such as LLaMA and Mistral to proprietary models, provided they are compatible with NVIDIA architecture.

Source: [NVIDIA Blog - RTX AI Garage Hermes Agent DGX Spark] (https://blogs.nvidia.com/blog/rtx-ai-garage-hermes-agent-dgx-spark/)