Language weights fine-tuned for typed decision questions (choice, yes/no and graded answers), packaged with the unchanged vision tower of the parent model. Submitted to the d3-lite soup pool.
Qwen/Qwen3.5-0.8B-Base