llama.cpp Release b12391 Adds Tool Call Support for Ministral-3
The llama.cpp project has released build b12391, which introduces support for tool calls within reasoning blocks specifically for Ministral-3 models. This update refines parser logic and handling of thinking end tags to ensure smooth integration of function calling features. This enhancement improves the capabilities of local LLM inference engines when dealing with advanced agentic workflows that require reasoning models to invoke external tools. Developers relying on llama.cpp for local deployment can now better utilize models like Ministral-3 for complex tasks requiring function execution. The pull request (#32271) addresses specific review feedback regarding Ministral-3's reasoning end tags by utilizing a choice between thinking termination tokens and tool call tokens. It also reverts unnecessary tag removals and includes updated tests to guarantee stability.
## BACKGROUND
llama.cpp is a widely used open-source library that performs local inference on large language models using the GGML tensor library. Tool calls allow language models to interact with external APIs or software functions by generating structured payloads during their generation process.