matrix07012 @matrix@gameliberty.club

Way too slow if you can't fit it in VRAM.
Letting the GPU do all the processing and letting it overflow into RAM is faster than having llama.cpp split it between CPU and GPU

32b 6k seem to work fine

and qwen3-vl heretic just stops thinking, wtf
{'role': 'assistant', 'reasoning_content': 'Okay, the user wants me to role', 'content': ''}

yep, gemma-12b heretic is too retarded

wtf, i don't think that's a bug, i think that's the model shitting the bed

Nope, it still made a bug, but wow, what a progress from gpt-3.5

Opus 4.5 can oneshot 600 line Mastodon bot

It's truly joever