Since the BC250 has limited memory, Qwen 35B-A3B is maxed out at IQ2M
The Uncensored model has issues with repeating the same output in coding and various other problems
I switched to gemma4 26B-A4B IQ3. Similarly, I used it with the heretic model
There's quite a difference
Qwen is so aggressively proactive that you'd wonder where its ethics are, like it's a terrorist or something
Gemma4 answers everything, but the tone is heavily subdued, giving off a "better not do it" kind of vibe
I tried adjusting MTP with limited resources, but Qwen had no effect. When increasing gemma4...
And in the code agent, gemma4's context consumption is much less
However, it doesn't seem to analyze code as deeply to that extent
