llama.cpp Release b11104 Adds Multi-Address Binding to llama-server
Release b11104 of llama.cpp introduces support for binding the llama-server HTTP component to multiple network IP addresses simultaneously. Implemented via PR #28690, this feature enables a single server process to accept connections across several network interfaces at once. This enhancement simplifies networking setup for self-hosted LLM deployment, allowing developers to expose llama-server to both local loopback and external local interfaces without running multiple instances. It offers greater deployment flexibility in mixed or containerized infrastructure environments. The update improves argument handling, adds explicit handling for overlapping network addresses, and intentionally rejects dynamic port selection (--port 0) when multiple TCP addresses are specified. Updated pre-built binaries have been released across various operating systems and backends, including CUDA 13.4, Vulkan, ROCm, and Snapdragon.
## BACKGROUND
llama.cpp is an open-source C/C++ library optimized for low-resource local inference of Large Language Models across CPU and GPU hardware backends. Its built-in llama-server utility exposes an OpenAI-compatible web API, allowing software applications to communicate directly with locally hosted AI models.