llama.cpp Release b11112 Adds Image Input Support in Server Function Call Outputs
llama.cpp build b11112 updates its built-in server component to support returning image inputs (`input_image`) within function call outputs. This allows multimodal models to receive visual data back from external tools executed during tool calling. This update enables developers to build complex multimodal AI agent workflows, allowing LLMs running locally or on dedicated servers to process images returned by external web scrapers or image processing scripts. It aligns llama.cpp's HTTP server API more closely with advanced multimodal tool-calling standards present in modern commercial APIs. The patch was implemented in pull requests #20663 and #22575, which also included minor code cleanups such as avoiding repeated type lookups in the server C++ codebase. The release includes pre-built binaries across multiple operating systems and hardware targets, including CUDA 12/13, Vulkan, ROCm, OpenVINO, SYCL, and Snapdragon NPUs.
## BACKGROUND
llama.cpp is a popular open-source C/C++ library designed for fast, memory-efficient inference of Large Language Models (LLMs) and vision-language models on local hardware. Function calling (or tool calling) is a technique where an LLM identifies when external code needs to be run, outputs structured parameters, and receives the tool's execution result back into its prompt context to formulate a response.