@NousResearch - Jensen Huang Ends GTC
@NousResearch on X Nvidia CEO Jensen Huang closed his GTC Berlin 2026 keynote with a Nous Research Hermes Agent demo in which the agent set up a simulated training environment for a Hugging Face Microduck robot. The post argues that open source lets users chain local and cloud models, devices, and tools together without waiting for a vendor to enable it.
- Huang installed and configured a customized Hermes agent with a single prompt, framing agent-building as asking an agent to build it.
- Hermes planned an obstacle course with Nemotron, and GPT-6 Astra built it using FreeCAD, Omniverse, Blender, Newton, and Isaac Lab.
- Hermes trained thousands of simulated ducks on the course until the policy was validated for the physical Microduck robot.
Jensen Huang ended his GTC Berlin 2026 keynote with a demo of Hermes Agent, the open-source agent from Nous Research, and the pitch was that you build agents by asking an agent to build them. He installed a customized Hermes from a single prompt, then pointed it at a real job: building a simulated training environment for the Hugging Face Microduck, a small robot duck. The Hermes instance was running on an Nvidia RTX Spark machine, and it used Nemotron to plan an obstacle course for the duck. GPT-6 Astra then did the construction work, drawing on FreeCAD, Omniverse libraries, Blender, Newton, and Isaac Lab.
The training step is where it gets concrete. Hermes ran thousands of simulated ducks on that course until it had a validated control policy, which is the piece that decides how the robot moves. Only then was the policy meant to go onto the physical robot. So the demo covered the full loop: scene design, physics simulation, reinforcement-style training across many parallel copies, and transfer to hardware. Whether the real-world transfer worked as cleanly as the keynote implied isn't something the post itself shows, so treat the finale as a demonstration of the pipeline rather than a benchmark result.
Nous's framing in the post is the more interesting part to me. Their argument is that open source matters less as a philosophy and more as a way to wire together whatever you need: local models, cloud models, devices, and tools, in chains you assemble yourself rather than waiting for a vendor to ship the integration. The Microduck demo is their example of that claim, with one agent coordinating several models and several design and simulation tools. Whether that kind of composability holds up outside a stage demo, with messier tools and less forgiving hardware, is the question the post leaves open.