Overview: It scored 2% (1.79% rounded up) It is 18/20th place scoring above Haiku 4.5 and Minimax M2.7 Full benchmark took 70 hours Average time per task 32m Average output tokens per task: 44k Perspectives: It scored suspiciously similar to 3.6 Plus and it really gets me wondering how the architecture of 3.6 Plus differs from 27B. Qwen 3.6 27B has a bad reputation in the community for being verbose. But surprisingly. The output tokens were on par or less to similar models....