Poor "intelligence" with Gemma4
I am testing Hermes with Gemma 4 (gemma-4-26B-A4B-it-OptiQ-4bit running with OMLX) on a Mac mini M4 64GB:
it is fast in terms of response times, but it sucks IMHO in terms of "intelligence" and "agency".
After several days spent (with the help of Claude) debugging and prompting I gave up: even a simple daily cron task (message me the output of curl -s 'https://wttr.in/{city1,city2,city3}?format=3') fails with hallucinations.
Same for a briefing (extracting news headlines from RSS feed links, for which I provided the python code as it was unable to code a working version of it): some times it works, most of the times it doesn't.
I was having much better results with much earlier explorations of OpenClaw.
Where am I wrong?