Looks like Step 3.7 Flash's long reasoning might get fixed ( llama.cpp )

https://github.com/ggml-org/llama.cpp/pull/25238

Turns out that trimming the input was the wrong thing to do.

Fingers crossed that this model can become useable soon. I'm still using Step 3.5 Flash because of how slow 3.7 has been in reasoning.