llamafile is a Mozilla Builders project, now being reworked and maintained by Mozilla.ai. Its core goal: to make open LLMs more accessible to both developers and end users. It achieves this by combining llama.cpp with Cosmopolitan Libc into a framework that condenses the complexity of LLMs into a single executable file (i.e., llamafile), which can run locally on most operating systems and CPU architectures, with no installation required.
Main capabilities and components:
- Single-file distribution: the model and runtime logic are packaged into one file, which can be executed directly to run an LLM locally.
- Cross-platform: supports most operating systems and CPU architectures, with no installation required.
- whisperfile: a single-file speech-to-text tool built on whisper.cpp using the same Cosmopolitan packaging approach, supporting transcription and translation of audio files, covering the same platforms, with no installation required.
- Versions and builds: since 0.10.0, a new build system has been adopted to make it easier to align with the latest llama.cpp, thereby supporting newer models and features, but some older features may be missing. Older versions are still available from the releases page. Prebuilt llamafiles display the bundled server version (e.g., 0.9.* examples, 0.10.* examples), making it easy to confirm the software version downloaded.
How to use:
- After downloading a prebuilt llamafile, run it directly; Windows users need to rename the file to add the .exe extension before running it.
- Windows note: only executables smaller than 4GB can run on Windows; llamafiles larger than 4GB cannot run on Windows. In this case, you can download the llamafile binary and run it with external weights/models (GGUF).
- Documentation entry point: docs.mozilla.ai/llamafile, covering prebuilt llamafiles, running llamafiles, creating llamafiles, installing from source, technical details, supported systems, troubleshooting, and more.
Use cases:
- Developers and end users who want to run open LLMs locally with zero installation.
- Scenarios that require distributing LLMs across different operating systems and CPU architectures.
- Scenarios that require local speech transcription and translation, where whisperfile can be used.
RestrictionsRestrictions and licensing:
- Executables larger than 4GB on Windows cannot be run directly and require using external weights/models instead.
- Starting with 0.10.0, while aligning with the latest llama.cpp, some features from older versions may be missing.
- Licensing: the llamafile project itself is Apache 2.0; modifications to llama.cpp and whisper.cpp use the MIT license to maintain compatibility and facilitate future upstream merging.
- The page logo was generated with the assistance of DALL·E 3.
Project address:https://github.com/Mozilla-Ocho/llamafile
