NIGHTRUN

NightRun is a local language model runtime that boots from a USB stick or SD card and runs without a conventional operating system. It is a UEFI application written in Rust that the computer's firmware starts directly. It copies a model entirely into RAM, then offers a chat on the screen framebuffer. There is no Linux, Windows or kernel underneath, and the runtime contains no network stack, so after the model loads nothing enters or leaves the machine. It supports Llama, Qwen and Granite models in quantized formats, with model files of roughly 1.3 to 2.4 GB checked by inline CRC 32 verification as they load. Inference runs on the CPU using hand written AVX2 and NEON kernels. The targets are x86_64 PCs booted from USB and the Raspberry Pi 5 booted from SD card, where the runtime even drives the board's fan. The chat interface draws its own terminal with streaming output, scrollback, line editing and commands such as clear and bye. The project states that greedy output is checked token for token against llama.cpp for every supported model family. An interactive installer builds the bootable image, downloads and verifies a model, and flashes removable media only after an exact typed confirmation. It is aimed at developers and tinkerers interested in offline, minimal footprint inference, and the vendor itself calls it weird software. The site publishes no price, and the project code is hosted on GitHub.