Reasoning seems broken with this release?
The non abliterated version correctly formats thoughts inside , but in the abliterated version it does not. I tested them both for about an hour. The abliterated verison goes on and on and on forever also.
Could you post the prompt and response you tested so we can see the differences?
Or you can send an email to [email protected].
Maybe you can post your prompt and response since yours works? Mine is just non stop failing to get out of think. I wasted a lot of time on my end because I thought my settings were wrong. I just used a super basic Prompt and only the base AI model. It's very simple to test. It fails to leave mode at the most basic responses. Tried both Q6 and Q8 (no Q6 for abliterated yet) versions. Once I used the pure version I realized my settings were fine all along since it never failed a single time.
Non abliterated - 100% chance to leave , but sometimes it does chop some off, but it will leave every time.
abliterated - Always fails to leave think
Also the models aren't even named yet. They have generic ggml names.
You can try overwriting the configuration file of Huihui-GLM-4.6V-Flash-abliterated with the original configuration file from GLM-4.6V-Flash, and regenerate the GGUF files again, except for the weight files.