llama.cpp NixOS flake breaks on recent commits; users pinning to three-week-old builds

llama.cpp's NixOS flake has stopped building on recent releases, forcing users to pin to commits from roughly three weeks ago to get working binaries. The issue surfaced this week, with users noting that attempts to build current releases fail consistently. For practitioners running NixOS — a declarative Linux distribution popular among developers who value reproducible builds — the breakage cuts off access to the latest llama.cpp features and optimizations.

GitHub issues for the project show a handful of build-failure reports, but they remain unconfirmed and have drawn little developer attention so far. It's unclear whether the breakage affects all NixOS environments or only specific machine configurations. NixOS users represent a small but vocal slice of the llama.cpp user base, and flake-specific build paths often lag behind the main CMake and Makefile workflows.

llama.cpp is the most widely deployed CPU and GPU inference engine for Llama, Mistral, Qwen, and other open-weight models. It ships optimizations for Apple Silicon, CUDA, ROCm, Vulkan, and SYCL backends, and the project typically sees dozens of commits per week. Users running NixOS who need llama.cpp inference are advised to pin their flake inputs to a commit from late April 2026 until the build path is fixed — a straightforward workaround for anyone already using Nix flakes, but one that leaves those users waiting while the rest of the ecosystem moves forward.