Creating complex software with zero coding experience with Qwen 3.8 27B run locally on a DGX Spark using Hermes harness.
The app is in production right now at the same time while I still develop it.
When the app cannot be reloaded because of important on going tasks there is a dev tree that is modified and tested before updating the production instance.
Qwen 27B decided on it's own to install "sensors" directly in our chat that notifies me and it of important events happening in the app.
I have never seen this before and I did not ask it to do it but it uses that to do real time debugging without me even needing to prompt it that something bad happen.
Actually, I don't even have time to see the problem and I see it already fixing it.
I don't even do the development in Hermes TUI or Hermes Desktop, this is straight Hermes agent in my messenger app on my phone/computer.
This stuff hits different.
Dude, when I checked my agent today after a power surge I found out that 27B installed sensors in our chat messages for a project it was working at.
I could not believe that this came from the model itself and taught @NousResearch added it in one of the updates and ask it to investigate if that came from the model or from a Hermes skill or updates and it reported that it came from the model.
No one asked it to do that and it did it just because I said something is important that was related to this.
I'll make soon a thread about this as I evolved the idea already much further (think JEV like model + GPIO).
Feel free to steal my idea !
Hope the guys from Hermes pick up on this message and they do it themselves too and if not I will publish something helpful for the community.
4
3
6
1,038
Another benefit of running Qwen 3.8 27B local on DGX Spark is that while doing decode the GPU power draw is just 40-44 W/h and that keeps the GPU under 65 C.
During prefill it will go 80+ W/h for short amounts of time.
Every number in this screenshot tells such and interesting story and it shows a complete view of how the model actually works in real world tasks inside a harness not in an artificial benchmark.
1
206
The projected API cost save info bubbles calculate using the last hour performance and the official Alibaba API pricing on OpenRouter.
Sep 25, 2026 路 2:13 PM UTC
1
75
Just to make an idea about the autonomy level of 27B developing in Hermes, I have 346 messages from the agent in the last 107 minutes.
These are mostly tool calls and reasoning traces but also messages it sent me.
This is why I like more the agent instead of the TUI because every decision it made it's documented and just one search away on my phone.
60