Any plan for Thinking version?
#1
by IlysvlVEizbr - opened
It would be nice if you could provide a REAP model based on Qwen/Qwen3-VL-30B-A3B-Thinking, and the corresponding GGUF file!
After reducing the number of experts, it does run on a 16G VRAM graphics card, but the accuracy of the document (image form) information extraction task drops significantly. I'd like to test if the accuracy improves to a usable level with the addition of Reasoning, so that I can actually do some real thing with a 16G graphics card.
I will consider this. :)
Currently, the REAP process only uses the Text dataset, so Experts that hold Vision information may have been deleted.