Ask for 5.0bpw and 6.0bpw quantization for GLM-4.5V-exl3 vision performance is sensitive to to quantization

[blackcat1402]

https://huggingface.co/turboderp/GLM-4.5V-exl3
I compared different quantization levels one by one and found this vision model is more sensitive than non-vision version GLM4.5 Air. Even with GLM4.5V 4.0bpw the text based coding quality is much lower that of GLM4.5 Air. I suspect the addition of vision function reduced its text output performance. However, the maximum quantization is 4.0 bpw, @turboderp @Could you provide 5.0bpw and 6.0bpw exl3 version so that I can find a "breakeven" point of GLM4.5V comparable to GLM4.5 Air. Thanks in advance.

[blackcat1402]

I apologize for being a complete novice when it comes to quantizing local models, but thankfully I've carefully read the exllamav3 documentation and am attempting some basic quantization work. Thank you for your excellent work; please ignore this request. There are some things I really need to learn on my own.