Piccolo-27B

AgentScope's (Alibaba) decision model built for tool review and context selection, with bidirectional attention and a reference implementation in-repo — the production-flavored end of the class, wired into the QwenPaw agent stack with a live Decision Playground. #14 on the Nov 6 draw (68.40); the card's own suite puts it at 84.4% on HarnessJevBench Hard and 82.6% on DataJevBench. Bilingual English/Chinese, Apache 2.0, ModelScope mirror.