Deconstruct @spydenator ·
Qwen3.8:27b has claimed its seat in my @NousResearch Hermes Agent as top orchestrator in my M4 mini 24GB machine.
"Hermes, gather the last 24 hours AI news and deliver it in Lumon industries micro-data refinement format mixed with Deconstruct Lab style and have it voice read it to me"
-- at a speed of ~16.8 tokens/sec generation, ~40.6 tokens/sec prompt-eval running locally on the mac mini - it took it 1 hour and 42 min to produce the daily refinement news.
next experiment, send the task mini mastered to my RTX-5090 machine and have it fly. Skill it up in Hermes and you got it producing daily morning news at the speed of light.
Have a listen 🔊
oh... and thank you, Kier -- I am his mouth, and through me, he will whisper on. 😅😉
What prompted it
The author wanted a themed daily AI news brief, read aloud, made by a model running locally.
What Hermes Agent did
- Set Qwen3.8:27b as the top orchestrator in Hermes Agent on the M4 Mac mini.
- Asked Hermes for the last 24 hours of AI news in the themed format, read aloud.
- It ran locally at about 16.8 tokens/sec generation and 40.6 tokens/sec prompt-eval.
- It took 1 hour and 42 minutes to produce the daily brief.
Evidence and limits
- Original post checked on X (Oct 19, 2026)
- Outcome reported by the author
- Not reproduced by anyone else yet
On the Mac mini it was slow: 1 hour and 42 minutes for one brief.
