Distilled Model's Vision Problem

Posted by TheAncientOnce@reddit | LocalLLaMA | View on Reddit | 2 comments

Have been using Qwen 3.6 Claude distilled version, 27b at Q4 for openclaw, Hermes and other local harnesses. But recently noticed that the Claude distilled version that I use lost its vision abilities. I'm serving the llm via lmstudio. Does anyone know if this is the case for all distilled claude?

[-]

a_beautiful_rhind@reddit

Catastrophic forgetting of understanding visual embedding due to the finetune. It's not the case with all or tied to claude outputs in particular. Try a different model till you find one that works or use a separate 2nd model as VLM.

[-]

TheAncientOnce@reddit (OP)

thanks. do you have a specific one that does work in mind?