A pull request in llama.cpp addresses an issue where trimming the input was incorrectly implemented, which previously hindered the performance of Step 3.7 Flash.
- The fix targets a specific bug in input handling that caused slow reasoning capabilities.
- This change aims to make the model usable for tasks requiring long-form reasoning.
The update is significant for users who have been avoiding Step 3.7 Flash due to its poor performance compared to earlier versions like Step 3.5 Flash.