i'm in cyber programs for both and with the newest models I can't do a lot of things (gpt5.5 is rather permissive, but the 5.6 announcement makes it sound like that's moving to corporate only). having exploits/RE in my history/memory for them both also makes me get blocked on things people wouldnt normally

Modified by: swiftraccoon on 2026-06-26 03:22:55 PM

Messing with https://ollama.com/huihui_ai/Qwen3.6-abliterated:35b-Claude-4.7
It sucks



Modified by: Mysticdrew on 2026-06-26 04:33:54 PM

It cannot even write files...

no idea about that model or person.

the appended claude 4.7 probably means it some time of finetune in addition. they don't mention how they abliterated it, etc.

I tried the non-claude 4.7 as well.

it was bad

They cannot write files

or edit files

definitely possible to mess them up. normal qwen3.6 is great at tool calls. i've not had to use abliterated for qwen3.6 though. it does whatever i want.

It refused to help me bypass eac

the abliterated just spat out a bunch of code in the terminal but would not write it to a file. So I've been trying to get it to write actual files and it keeps saying it is written but no file exists.. lol

it's all in the phrasing.

"create and write to disk an empty text file in g:\test"

not referring to the abliterated's ability to do tool calling