MiMo-V2.6-Flash on vLLM: fixes for "empty responses" with thinking + tools, and a hidden 2,048-token output cap

A user addressed issues with MiMo-V2.6-Flash on vLLM, specifically "empty responses" when using thinking and tools, and a hidden 2,048-token output cap. The problems were identified as serving bugs rather than model deficiencies. One key finding was the lack of reasoning effort control; while Xiaomi's API and vLLM recipes only offer thinking on/off, an experimental "think thoroughly" hint significantly increased thinking tokens and time, improving correctness. These fixes resolved multi-step tool tasks with thinking enabled.