How usable are "lobotomized" weights?
retained text
This subreddit talks a lot about PP/decode speeds but I feel like I hear comparatively less about actual real-world usability. I was playing around with Strata hosting QFN3.8 iq3_s and Qwen3.6 35B Q8 over the weekend and despite being the more hyped, I found that QFN was less "usable" as a coding agent than the smaller Qwen3.6. Even when using the recommended temp, top-k, etc. settings that I hear supposedly stop the overthink and doomloops, I still couldn't get QFN to be reliable. Qwen3.6 on the other hand, had a very narrow interpretation of my prompts and we would have to go back and forth to clarify edge cases, but it ultimately would execute what I asked it to. If I had to use an analogy, QFN iq3 feels it has dunning krugers with illusions of grandeur while Qwen3.6 felt like a talented but unmotivated programmer. If given the choice, I'd rather work with the capable guy who hates his job and wants to FIRE than the guy who talks a big game but can't deliver. Obviously QFN iq3 is "lobotomized" while the Qwen3.6 Q8 is running at a far higher precision, but I was surprised that there's so much buzz about QFN iq3 when I was having such a comparatively worse experience with it. Is there something I'm missing and should be doing to improve QFN (aside from buying more hardware obviously)? This experience with QFN has made me wonder if Qwen 4 is going even going to be meaningful for me considering how getting the mid-size MoEs to even load will require lobotomization as well.
document fetched 2026-11-03 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xmb2qh/how_usable_are_lobotomized_weights