Key Takeaways
- Apache 2.0 license; the project's own changes to llama.cpp and whisper.cpp remain MIT-licensed to stay upstreamable
- Roughly 25,900+ GitHub stars as of September 2026, repository at github.com/mozilla-ai/llamafile
- Created by Justine Tunney, creator of Cosmopolitan Libc, first released by Mozilla's innovation group in November 2023
- One executable file runs on macOS, Linux, BSD, and Windows β Windows specifically caps executables at 4 GB
- GPU acceleration available via Metal (Apple Silicon), CUDA (NVIDIA), ROCm (AMD), and Vulkan
- Built directly on llama.cpp for the actual model inference, with Cosmopolitan Libc solving cross-platform packaging
π In One Sentence
llamafile is a free, Apache 2.0 project, created by Justine Tunney and originally released by Mozilla, that packages a model and the llama.cpp engine into a single executable file that runs unmodified across macOS, Linux, Windows, and BSD using Cosmopolitan Libc.
π¬ In Plain Terms
Download one file and double-click it β no installer, no picking a platform-specific build β because the file itself is valid on several different operating systems at once, a trick made possible by Cosmopolitan Libc rather than by llamafile packaging multiple separate builds together.
πNote: Version 0.10.0 introduced a rebuilt build system to keep llamafile aligned with current llama.cpp versions, and CUDA support for Linux was reintroduced in February 2026 after a period of being out of sync β check the release notes for the exact feature set of the version being downloaded.
What Is llamafile?
llamafile is a free, open-source project that packages a language model together with the llama.cpp inference engine into a single executable file, distributable and runnable without installation on multiple operating systems. It was created by Justine Tunney and first released by Mozilla's innovation group in November 2023.
The project is hosted at github.com/mozilla-ai/llamafile (previously under the Mozilla-Ocho organization) and has passed roughly 25,900 GitHub stars. It is released under the Apache 2.0 license, with the project's own modifications to llama.cpp and whisper.cpp kept under MIT to remain compatible with, and upstreamable to, those projects.
- Combines a model file and the llama.cpp engine into one distributable executable, rather than requiring a separate engine install plus a separate model download
- Built on top of llama.cpp for the actual inference work β llamafile's own contribution is the packaging and cross-platform execution layer
- Runs as a chat web UI by default when launched, alongside an OpenAI-API-compatible server mode
- Prebuilt example llamafiles for popular open-weight models are published on Hugging Face by the project and community members
What Is Cosmopolitan Libc, and How Does It Make This Possible?
Cosmopolitan Libc is a C standard library, also created by Justine Tunney, engineered to compile "fat" binaries that are simultaneously valid executables on multiple operating systems and CPU architectures. This is the underlying technology that makes llamafile's single-file, cross-platform distribution possible.
- A binary built with Cosmopolitan Libc can run, unmodified, on macOS, Linux, BSD, and Windows, rather than requiring a separate compiled build per operating system
- This differs from packaging multiple platform-specific binaries into one archive β it is genuinely one binary that is valid across systems
- Windows imposes a 4 GB file-size limit on executables, which caps how large a single llamafile can be specifically on that platform
- Cosmopolitan Libc is a separate, general-purpose project; llamafile is one prominent application of it, not the only one
How Do You Download and Run a llamafile?
Download a prebuilt llamafile for the model you want, make it executable, and run it directly β there is no separate installation step. Building a custom llamafile from your own GGUF model file is also supported.
- 1Download a prebuilt llamafile (a model bundled with the engine) from the project's Hugging Face page or the GitHub Releases page.
- 2On macOS and Linux, mark the file executable (
chmod +x yourmodel.llamafile) if it is not already, then run it directly (./yourmodel.llamafile). - 3On Windows, rename the file to add a
.exeextension if it does not already have one, due to how Windows identifies executables. - 4By default, running the file launches a local chat web UI in your browser, alongside an OpenAI-API-compatible server endpoint.
- 5To enable GPU acceleration, pass
-ngl 999to offload as many layers as possible to a detected GPU (Metal, CUDA, ROCm, or Vulkan). - 6To build a custom llamafile from your own GGUF model, use the project's packaging tools to combine the model with the llamafile engine binary.
Does llamafile require installation?
No. Download the file, make it executable if needed, and run it directly β there is no separate installer.
Can llamafile run without a GPU?
Yes. CPU-only operation works out of the box; GPU acceleration (Metal, CUDA, ROCm, Vulkan) is an optional performance improvement, not a requirement.
What Hardware and GPU Acceleration Does llamafile Support?
llamafile supports CPU-only operation out of the box, with optional GPU acceleration across most major vendors. GPU support has evolved over time, so the exact capability depends on the version downloaded.
- Apple Silicon β Metal acceleration, enabled by default when a Metal GPU is detected (implemented December 2025)
- NVIDIA β CUDA acceleration on Linux, reintroduced in February 2026 after a period of being out of sync with upstream llama.cpp
- AMD β ROCm acceleration
- Cross-vendor β Vulkan backend support
- GPU libraries (CUDA, ROCm, Vulkan) are dynamically loaded from prebuilt files alongside the executable rather than compiled permanently into it
Who Should Use llamafile?
Use llamafile if distributing or running a model as a single, dependency-free, cross-platform file matters more than having the newest llama.cpp features immediately; use llama.cpp directly or a wrapper app if that single-file portability isn't the priority.
How Does llamafile Compare to KoboldCpp and Other Zero-Install Tools?
llamafile's closest comparison is KoboldCpp β both distribute as a single file built on llama.cpp β but they solve different problems: cross-platform portability versus a bundled roleplay-focused UI.
Common Mistakes When Evaluating llamafile
Most confusion comes from misunderstanding how the single-file cross-platform trick works, or expecting llama.cpp's newest features immediately.
Frequently Asked Questions
What is llamafile?
llamafile is a free, Apache 2.0 project that packages a model and the llama.cpp engine into a single executable file that runs unmodified across macOS, Linux, Windows, and BSD, using Cosmopolitan Libc.
Who created llamafile?
Justine Tunney, creator of Cosmopolitan Libc, created llamafile, first released by Mozilla's innovation group in November 2023. It is now maintained at github.com/mozilla-ai/llamafile.
How can one file run on macOS, Linux, and Windows?
llamafile is built with Cosmopolitan Libc, a C standard library designed to produce a single binary that is a valid executable on multiple operating systems and CPU architectures at once, rather than requiring separate builds per platform.
Is llamafile free to use commercially?
Yes. llamafile is Apache 2.0-licensed, free for personal and commercial use. Its own changes to llama.cpp and whisper.cpp remain MIT-licensed.
Does llamafile require a GPU?
No. CPU-only operation works by default; GPU acceleration via Metal, CUDA, ROCm, or Vulkan is optional and improves speed.
Is there a file-size limit for a llamafile?
Yes, on Windows specifically β Windows caps executable files at 4 GB, which can limit how large a model can be packaged as a single llamafile for Windows users.
Does llamafile stay up to date with llama.cpp?
It syncs periodically rather than continuously β version 0.10.0 introduced a rebuilt build system specifically to keep pace with current llama.cpp versions, and CUDA support for Linux was reintroduced in February 2026 after a period of lagging behind.
What is the difference between llamafile and KoboldCpp?
Both distribute as a single file built on llama.cpp with no separate installer, but llamafile's distinguishing feature is genuine cross-platform portability (one file runs on 4 different operating systems), while KoboldCpp's distinguishing feature is a bundled roleplay- and story-writing-focused web UI.
Can I make a llamafile from my own model?
Yes. The project provides packaging tools to combine a GGUF model file with the llamafile engine binary into a custom single-file executable.
Is llamafile good for production multi-user serving?
Not its focus. llamafile is built for simple, portable single-file distribution and single-user or small-scale use; for high-throughput multi-user production serving, vLLM or SGLang are the more appropriate tools.