Mozilla.aillamafile packages an open LLM's weights together with its inference runtime into one executable that runs locally on most operating systems and CPU architectures without installation. It combines llama.cpp with Cosmopolitan Libc so that the same file works across Linux, macOS, Windows, and BSD; you download a .llamafile, mark it executable (or add .exe on Windows), and run it to get a local chat server. The llamafile binary can also be run on its own against external GGUF weights, which is the route for models over 4 GB on Windows where executables above that size do not run.
ai. Versions from 0.10.0 onward use a new build system that tracks upstream llama.cpp more closely, so they support newer models and functionality while some features of earlier releases may be absent; the 0.9.x line remains available for those. Pre-built llamafiles on Hugging Face indicate which server version they bundle.
whisperfile applies the same single-file packaging to whisper.cpp, giving a speech-to-text tool that transcribes and translates audio files across the same platforms with no installation.
Features
- Single-file distribution: model weights and runtime ship as one executable built on llama.cpp and Cosmopolitan Libc
- Cross-platform execution: one file runs on Linux, macOS, Windows, and BSD across CPU architectures
- External weights: the standalone llamafile binary loads any GGUF model, including those above the 4 GB Windows executable limit
- Pre-built llamafiles: ready-to-run models such as Qwen3.5 0.8B are published on Hugging Face with the bundled server version labeled
- Creating llamafiles: documentation covers packaging your own model into a llamafile and building from source
- GPU support: documented paths for using GPUs with both llamafile and whisperfile
- whisperfile: single-file speech-to-text based on whisper.cpp with transcription, translation, and a server mode
- Version lines: 0.10.x tracks current llama.cpp; 0.9.x keeps the earlier feature set and remains downloadable
