Brain damage?
I tried this with the same settings I used with the regular qwen3.5 35b a3b and it is outputting text correctly, makes sentences, but I think it has suffered catastrophic brain damage. This one drools, at least for me. Anyone else getting a better experience?
It's still early in my testing, but it seemed functional for initial coding tasks. It'll be interesting to have a more thorough evaluation though.
"heretic" not working as it should, compared to the older Q3 models. It's probably still too early...
This model works better than any other heretic/ablated finetune. In my experience, using Q4 from @mradermacher , it produces code accurately, can handle controversial discussions, and even NSFW scenarios, flawlessly.
It has beaten beaten qwen2.5 coder in code (full gui python base64 encode/decode, self contained app, all to all file type, with full os/sys access, preview window etc.) all this in 6 seconds, with zero errors.
So for anyone having trouble with it, it must be your settings, for sure.
My honest review is that it does more than it's supposed to be able to do, per the original training. It does all we need, and want. Which is great.