Board:   [DoveNet] Internet Discussion
Subject: AI LLM Artificial intelli
From:    phigan
Date:    Fri Oct 02, 2026 06:47

Subject: AI LLM Artificial intelligence infrastructure
@VIA: TACOPRON
  Re: AI LLM Artificial intelligence infrastructure
  By:bbsing to phigan on Sun Sep 27 2026 08:08 pm

 > I've been using Hermes with llama backend on a Strix Halo hardware, and
 > Qwen3.6-35B-A3B-Q8_0.

Totally agree with everything you said in your post. BUT, try Qwen3.8:27b or
GLM-anything. Try oh-my-pi (https://omp.sh) for programming in place of
Hermes. Use Hermes for tasks.

Not sure how much (v?)RAM your hardware has.. I've got 24gb and had to switch
from Ollama to use qwen3.8:27b, even though 3.6:27b and 3.6:35b (same one
you've got) ran fine. 3.8, even smaller ones, would just stop mid thinking on
the very first turn. I'm using some build off github..

github.com/MiaAI-Lab has arepo for it that says 16gb nvidia gpus

github.com/syv-ai/HyperQwen is an interesting looking one but it didn't work
for me immediately and I havem't gotten back to it to find out why.

Oh, also try to keep "reasoning" on low or 'none'. Try GLM (whatever you can
find that fits) whike you're still on Ollama, too. I got some
4.7-flash-ablitterated and then some really tiny 5.3 that was impressive for
how small it was.

---
 ■ Synchronet ■ TIRED of waiting 2 hours for a taco? GO TO TACOPRONTO.bbs.io
.