please provide vision mmproj file

#4
by akierum - opened

What version of vision mmproj file works best for this model?

protoLabsAI org
edited Jul 9

ThinkingCap is a text-only finetune of Qwen3.6-27B, so its vision tower is identical to the base; any correct Qwen3.6-27B projector works.
I've just added one converted straight from the base weights: mmproj-ThinkingCap-Qwen3.6-27B-f16.gguf . Verified it against the trunk and it reads images and OCRs text fine.

Thanks, why ulsloth includes mmproj-F32.gguf with Q8 normal and MPT versions downloaded with LMstudio?

protoLabsAI org

I'm not sure, perhaps its because the official fine tune release includes the mmproj? I've only quantized this model, I wasn't the original author. https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.6-27B-GGUF

Your version is way worse on visual tasks then the Unsloth version, is it because of mmproj file, or the thinks less means quality is also lower despite benchmarks being the same?
Menu recreation from site https://iamkate.com/code/tree-views/

compareqwen

artificial-citizen changed discussion status to closed

Sign up or log in to comment