Vision + coding + reasoning merge. Qwen2.5-VL language backbone used as base,
with Qwen2.5-Coder-7B-Instruct (0.6) and DeepSeek-R1-Distill-Qwen-7B (0.4)
merged via DARE-TIES. Original VL vision tower reattached.
Run with llama.cpp / LM Studio
Use Qwen-VL-Coder-R1-7B-Q4_K_M.gguf together with
mmproj-Qwen-VL-Coder-R1-7B-f16.gguf for image support. Fits in 8GB.
