Not a VLM
Hi,
Thanks a lot for releasing this modification of Qwen, seems a great version to use to play around and debug things. One thing I noticed is that, despite your mentioning that the vision encoder weights are conserved, the model cannot actually load as a VLM. I think it needs some adjustments to the config,json and it requires an import of the processor configs.
Just wanted to let you know in case you'd like to update the repo :)
For now, I'll use alphaedge-ai/Qwen3.5-4B-eng-32768, although it's a bit bigger. If I find the time, I may try adding the required files to this repo myself.
Best,
Daniel
Thanks for catching this — you were right,. The model was built by loading the base model via AutoModelForCausalLM, which silently loads only the inner text decoder and drops the vision tower entirely, even though the card claimed vision was preserved. It wasn't a missing config field; the vision weights were never in the checkpoint at all.
Rebuilt it properly from the full Qwen3_5ForConditionalGeneration class this time (vision tower + multimodal wiring genuinely intact), verified with a real image+text generation pass in both transformers and mlx_vlm, and pushed the fix. Should load correctly now. Appreciate the report.
Thanks again!