Just wanted to say “Hello” and thank you for all the insights you are sharing here.
After following your Board for some time, I hope I can contribute to it in future.
Maybe someone is interested in my Setup:
- OpenClaw running on an isolated AMD Strix-Halo 128GB - with main agent: qwen3.5:35B - fp8
- That main agent is happily spawning Sub-Agents, that run on my DGX Spark or a RTX5090
The main benefits are:
- OpenClaw stays responsive, while the heavier tasks are performed on the DGX Spark (ollama with Nemotron3-super). For smaller, high speed tasks the 5090 steps in. That system is completey local and running stable since approx. 30 days now.
- The Strix Halo is running in an isolated VLAN behind its own router. That keeps my DGX Spark safe to work without being concerned about safety issues from OpenClaw. I had a look at NemoClaw, which looks interesting. Still I personally would not like to run OpenClaw/ Nemo Claw in an isolated sandbox. Just to much hassle with the sandbox settings and for the time being too risky in case you configure something wrong.
I am hoping to get the NVfp4 Nemotron models running on DGX Spark soon. I tried nemotron-3-nano:30b in NVFP4 the other week with the avarok build. It was running lightning fast, but kept crashing every 10min. So Nemotron is now running as GGUF, which is okay for now.