Edge and embodied systems increasingly rely on vision‑language models (VLMs) to handle heterogeneous skills, yet the common practice of attaching a separate task adapter to each skill inflates memory usage and management complexity. UniCompress introduces a deployment‑oriented framework that folds multi‑task expertise into a single adapter‑free compact model.\n\nThe process starts from a consolidated BF16 Qwen3‑VL‑4B teacher. An asymmetric INT2 student is built using GPTQ initialization, followed by teacher‑guided quantization recovery that jointly quantizes and restores the projection weights in both the vision encoder and the language backbone. This joint treatment enables aggressive INT2 compression of the entire vision‑language network while preserving its multi‑task behavior.\n\nEvaluation spans four heterogeneous embodied tasks: posture recognition, household‑safety understanding, multimodal dialogue routing, and mobile‑GUI action prediction. Weight storage shrinks from roughly 8.9 GB in BF16 to 2.1 GB in INT2—a reduction of about 76%. The INT2 macro‑task score improves by 34.3 percentage points over plain GPTQ‑INT2 and stays within 0.9 points of the BF16 teacher.\n\nThe results demonstrate that teacher‑guided joint vision‑language INT2 recovery can deliver a single multi‑skill VLM on resource‑constrained platforms, offering substantial storage savings without sacrificing performance.\n\nReview: UniCompress achieves unprecedented compression by co‑quantizing cross‑modal weights, providing a practical pathway for deploying a unified multi‑task VLM in edge intelligence scenarios.