mozilla-ai/llamafile

What it does

llamafile packages llama.cpp with Cosmopolitan Libc so that a model and its inference runtime live in one executable file. Distribution becomes a file download rather than an install process, and the same file is intended to run locally across common operating systems and CPU architectures. The project is maintained under Mozilla.ai after starting as a Mozilla Builders project (source:21).

A companion tool, whisperfile, applies the same single-file packaging to speech-to-text with whisper.cpp, covering transcription and translation of audio files.

It is aimed at two groups: developers who want a simple way to ship an open model to users, and end users who want to try a local LLM without setting up Python, CUDA or a container stack.

Quick start

The README gives a minimal macOS/Linux/BSD flow using the small Qwen3.5 0.8B example (source:33):

curl -LO https://huggingface.co/mozilla-ai/llamafile_0.10/resolve/main/Qwen3.5-0.8B-Q8_0.llamafile
chmod +x Qwen3.5-0.8B-Q8_0.llamafile
./Qwen3.5-0.8B-Q8_0.llamafile

On Windows, rename the file with an .exe extension before running it. This example is the smallest prebuilt llamafile, chosen because it is the most likely to work on modest hardware; larger models are listed in the pre-built llamafile catalog. The quickstart documentation covers additional paths, including running the llamafile binary against external GGUF weights.

Constraints

Windows only runs executable llamafiles under 4 GB, so larger files must use the binary plus external GGUF weights instead (source:35). Version 0.10.x uses a new build system aligned with newer llama.cpp releases, and some features from the older releases may be missing; the previous "classic" versions remain available from releases.

The project is Apache 2.0 licensed, while the changes to llama.cpp and whisper.cpp remain MIT so they can stay upstreamable. Supported platforms are detailed in the support reference.

HotRepo's take

In the HotRepo daily C++ ranking for 2026-11-10, llamafile recorded 18 new stars and reached position 10, with a total of 26,602 stars and 1,675 forks at that snapshot — solid but not explosive for a project of this size. A December 2025 Hacker News discussion (43 points, 7 comments) surfaced two recurring concerns: GPU acceleration is limited, with CPU inference as the main optimization target as one commenter noted, and another commenter questioned whether the project was actively maintained after a May release with few visible commits asked publicly. Adoption for convenience-oriented use — "one file, point it at a .gguf" — is the clearest signal in that thread, and there is even a community transfer tool built for moving multi-GB llamafiles between machines mentioned alongside it. Readers deciding on llamafile should weigh the simple distribution model against the GPU and maintenance questions raised there.