llama.cpp Release b11057 Adds Dedicated Chat Parser for Ling 3.0 (Bailing V3)
Release b11057 of llama.cpp introduces a specialized chat parser for the Ling 3.0 (Bailing V3) model family to correctly isolate tool calls occurring within reasoning blocks. The new parser terminates reasoning output when encountering either a closing `</think>` tag or the start of a `<tool_call>` tag. Previously, Ling 3.0 models running on llama.cpp failed to execute tool calls because unclosed reasoning blocks caused tool calls to be misclassified as pure text output, breaking agent execution loops. Aligning llama.cpp's parser logic with frameworks like vLLM and SGLang restores proper agentic function-calling capabilities for Ling 3.0. Ling 3.0 Flash templates pre-open `<think>` tags in generation prompts without emitting an explicit opening tag, allowing `<tool_call>` to arrive before `</think>`. Detection for this specialized parser is safely gated using unique `<role>...</role>` section markers unique to the Ling 3.0 template family.
## BACKGROUND
llama.cpp is a popular open-source LLM inference engine written in C/C++ for executing models efficiently across diverse hardware. Modern reasoning LLMs use structured tags like `<think>` and `<tool_call>` to separate internal thinking steps from function calls, requiring dedicated parsers to route generated tokens correctly to agent applications.