Editorial illustration for Hermes deploys self‑improving AI agents using NVIDIA RTX PCs and DGX Spark
Hermes deploys self‑improving AI agents using NVIDIA RTX...
Self-improving AI has become a business goal, a bullet point on investor decks. Hermes says it’s now a working product. The company is deploying autonomous agents that learn and adapt in real time.
They claim it’s not a simulation. It’s running on a specific, fairly common hardware stack: NVIDIA RTX PCs handle distributed tasks, and a compact unit called the DGX Spark sits at the center.
That little box is the key. It has 128GB of unified memory and a petaflop of AI performance. It’s designed to run large models, specifically 120-billion-parameter ones, continuously.
A newer, smaller model called Qwen 3.6 35B is also in the mix, speeding up concurrent jobs. This isn't about a single smart program. It's a system built to let many agents work and learn all day without falling over.
NVIDIA DGX Spark is the ideal companion -- a compact, efficient standalone machine built for sustained, all-day agentic workflows. With 128GB of unified memory and 1 petaflop of AI performance, NVIDIA DGX Spark can run 120 billion-parameter mixture-of-experts models all day. And the new Qwen 3.6 35B model delivers equivalent intelligence in a leaner footprint -- running faster and giving users the capacity to run concurrent workloads.To maximize performance and ease of use, read the nvidia.com/spark" target="_blank">Hermes DGX Spark playbook. Plus, register for upcoming hands-on sessions in NVIDIA's "Build It Yourself" agentic AI series to learn how to build autonomous AI agents with NemoClaw and OpenShell.NVIDIA DGX Spark is available to order from NVIDIA's manufacturing partners -- visit the marketplace.
The pitch is straightforward. The hardware exists. The models are getting more efficient.
Hermes argues this specific architecture of Spark at the core and RTX at the edge is what makes self-iterating agents possible outside a lab. They've published a playbook. NVIDIA is running workshops.
The units are for sale. The implication is that the barrier is no longer technical, but a matter of deciding to build. Whether that's true depends on your tolerance for vendor promises and the real cost of running a petaflop machine all day.
Related Reading
AI Tools & Apps Claude gains shared context in Excel, PowerPoint; Microsoft adds Copilot Cowork
AI Tools & Apps Windows Copilot AI unable to pinpoint image source in user test
AI Tools & Apps LG's recent webOS update adds Microsoft Copilot app, now removable
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
Related Topic NVIDIA NeMo powers telco reasoning model for autonomous network workflows
Industry Applications
From Archives WhatsApp launches Meta AI Incognito Chat, cuts latency for privacy
From Archives AutoScout24 cuts development cycles to days using AI‑powered Codex
From Archives NVIDIA and SAP introduce NemoClaw blueprint to add trust to specialized agents
Open Source
From Archives NVIDIA's Star Elastic bundles 30B, 23B, 12B models; 23B hits 85.63 on AIME-2025
LLMs & Generative AI
Common Questions Answered
What hardware stack does Hermes use to deploy self-improving AI agents?
Hermes deploys autonomous agents using NVIDIA RTX PCs to handle distributed tasks, with a compact unit called the DGX Spark serving as the central hub. The DGX Spark features 128GB of unified memory and delivers a petaflop of AI performance, making it the critical component that enables self-iterating agents to function outside laboratory environments.
How does the DGX Spark enable real-time learning in Hermes' autonomous agents?
The DGX Spark's 128GB of unified memory and petaflop-scale AI performance allow it to process and adapt agent behavior in real time without requiring cloud infrastructure or simulation environments. This architecture enables the autonomous agents to learn and improve continuously during actual deployment rather than in controlled lab settings.
What makes Hermes' self-improving AI agents different from simulations?
Hermes emphasizes that their self-improving AI agents are not simulations but working products running on actual hardware in production environments. The company claims this distinction is significant because the agents operate on a specific, commercially available hardware stack of RTX PCs and DGX Spark units, demonstrating practical viability beyond theoretical models.
Why does Hermes argue the technical barrier for self-iterating agents has been removed?
Hermes contends that with efficient AI models, published playbooks, NVIDIA workshops, and commercially available DGX Spark units, the technical obstacles to deploying self-iterating agents have been overcome. According to the company, the remaining barrier is no longer technical capability but rather organizational decision-making about whether to implement this architecture.
Further Reading
- Hermes Unlocks Self-Improving AI Agents, Powered by NVIDIA RTX PCs and DGX Spark — NVIDIA Blog
- Run Hermes Agent with Local Models | DGX Spark — NVIDIA Build
- New Playbook: Run Hermes Agent with Local Models — NVIDIA Developer Forums
- Hermes Agent and Qwen 3.6: Local AI Supercharged by NVIDIA RTX and DGX Spark — EJS Computers Blog
- 2026 - Hermes Unlocks Self-Improving AI Agents, Powered by NVIDIA RTX PCs and DGX Spark — Tacktech
