[Open-Source Release] AA-VL-4B: A 4.4B GUI Grounding VLM trained on ~7K curated rows ($8 compute) achieving 65.25% / 74.47% on OSWorld-G (Apache-2.0, GGUF + full per-sample eval logs included)

AA-VL-4B is an open-weight 4.44B GUI grounding vision-language model (r=32 LoRA merged into Qwen/Qwen3-VL-4B-Instruct, Apache-2.0) designed for computer-use agents. Given a desktop/web screenshot and a natural-language instruction, it outputs a single normalized click coordinate (click x= y= on a 0…