Multimodal LLM (Qwen3-VL-30B-A3B-Instruct, MoE 3B active). Best for text reasoning, OCR, structured extraction. 32K context. NOTE: deployment currently in text-only mode — image_url parts are accepted but not interpreted.