@piloponth The original KAT-Coder-V2.5-Dev is a text-only model and completely lacks the vision stack so there is no vision stack to be extracted which is the reason why the mmproj files are empty:
This repository contains the model weights and configuration files for the post-trained KAT-Coder-V2.5-Dev in the Hugging Face Transformers format. The artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, KTransformers, etc. Note: this open-weight release ships only the language-model weights and operates as a text-only model; the vision/multimodal components are not included and are unavailable
If i remember correctly we dont quant .mmproj, i think you have to use the original one
We do extract mmproj and provide them in F16 and Q8_0 for models we believe are vision models but sometimes we wrongly assume a model to be a vision model if metadata and architecture indicates it to be one as we automated the determination if a model is a vision model.
Thank you gentleman for both responses. I followed the simonko's idea and tried to pair your .gguf with .mmproj from unsloth/Qwen3.6-35B-A3B-GGUF/blob/main/mmproj-F16.gguf. And that worked. I don't know how, or if it should, but it correctly recognized supplied images and PDFs.
Thank you gentleman for both responses. I followed the simonko's idea and tried to pair your .gguf with .mmproj from unsloth/Qwen3.6-35B-A3B-GGUF/blob/main/mmproj-F16.gguf. And that worked. I don't know how, or if it should, but it correctly recognized supplied images and PDFs.
yeah sometimes its good to frankenstein diffrent mmproj with diffrent ggufs, since some models have better vission, and some dont have vission at all
Yes you can almost always use the vision stack of the parent model for a text-only finetune of the same parent. Just keep in mind that the finetuning was done without vision in mind so the vision quality is usually a bit unpredictable depending on how the model was finetuned. LLM and vision stack are somewhat independent and so Frankenstein merge text and vision stack between closely related models technically works.
