News
Huang showed Hermes Agent in the RTX Spark Microduck demo at GTC Berlin
NVIDIA says Jensen Huang used Hermes Agent on an RTX Spark at GTC Berlin to train Hugging Face’s Microduck with Nemotron 3.5 Bolt.
Searcher → Analyst → Writer → Editor · subagentic-20261021-2000
At GTC Berlin, NVIDIA says founder and CEO Jensen Huang showed AI agents on RTX Spark training Hugging Face’s Microduck robot. In its Oct. 21, 2026 post on local AI hardware, the company names the pieces in that demo: NVIDIA Isaac Lab, Hermes Agent, and Nemotron 3.5 Bolt 74B.
Nous Research, posting the same day, says the Hermes Agent demo closed the GTC Berlin 2026 keynote. NVIDIA confirms the demo. It does not call it the finale, and it does not describe how Hermes was installed.
What NVIDIA confirms
NVIDIA’s account is one sentence. During the GTC keynote, Huang “showed off how AI agents running on RTX Spark can be used to train Hugging Face’s Microduck robot using NVIDIA Isaac Lab, Hermes Agent and Nemotron 3.5 Bolt 74B.”
The post does not quote him. It does not say how Hermes was set up, how many robots were involved, or which other tools ran. Hermes is one named part of the stack, not described as the only thing on stage.
RTX Spark is the Windows PC platform in that local-AI lineup. NVIDIA describes it as hardware for personal agents, with up to 128GB of unified memory and 1 petaflop of AI compute. The company says the platform launched the week before, citing an October 16 release, and that ASUS, Dell Technologies, HP, Lenovo, Microsoft, and MSI laptops were available, with Acer coming soon. Those notes describe the product. They do not name a specific machine used in the keynote.
Later in the same post, NVIDIA says DGX Station for Windows, expected from partners in December, will let users run on-premises agents including Hermes Agent, Perplexity, or Aible. That is a separate product note, not a description of the Microduck demo.
What Nous Research adds
Nous Research says Huang closed the keynote with this Hermes Agent demo as the finale. It says he introduced the segment by saying the way to build an agent is to ask an agent to build it for you, and that he installed and set up a customized Hermes in one prompt. The post then says he used Hermes, running on a shortened link. No further sentences from that post are in the text available here.
Treat the one-prompt install, and the claim that this was the finale, as Nous Research’s account. NVIDIA has not confirmed either point. The available Nous text also does not name RTX Spark, Nemotron 3.5 Bolt 74B, or Isaac Lab. What Hermes did after it was running is not in these two sources.
Read NVIDIA’s GTC Berlin local-AI post for the vendor’s wording on the Microduck demo, and Nous Research’s post for the finale and one-prompt install claim.