ggerganov/llama.cpp released b11951
Release b11951 of llama.cpp introduces fixes for stripping trailing whitespace from parsed reasoning and content, and properly stops Qwen3-coder reasoning before template separators. This update improves the accuracy and formatting of chat session outputs when running models like Qwen3-coder locally, ensuring cleaner interaction logs and proper token handling. The release includes improvements for handling partial prefixes until a match at the end of generation, along with updates to chat parsing tests.
## BACKGROUND
llama.cpp is an open-source software library that enables efficient inference for various large language models using the GGML tensor library. It serves as a core backbone for numerous local AI execution tools across multiple hardware platforms.