25.7k stars · Apache-2.0 · 0.10.6, read from /releases/latest, shown as '15 Sep' with no year and therefore 15 September 2026. Code pushed 21 August 2026, proved via ungh.cc · Track this in Scout
It packs a language model and the program that runs it into one executable file that works on macOS, Linux and Windows without being installed.
▶Repo detailsthe review · specs · pros & cons · install
What it is
A project from Mozilla that packs llama.cpp and a model into a single executable file that runs on macOS, Linux and Windows without being installed. It also ships whisperfile, which does the same trick for speech-to-text.What it is good for. Anyone who wants to try a model without accounts, keys or bills, and anyone who cannot send text to another company. A lawyer, a doctor and a founder with an unreleased product all have the same problem, and this is the shortest answer to it. It is also the easiest way to hand someone a working demonstration: one file, one command.
- Apache-2.0, and the file runs on three operating systems without being installed or configured.
- Nothing you type leaves the machine, so private text stays private.
- It works without a graphics card, though a graphics card makes it much faster.
- Windows will only run a file under 4 GB, so a larger model needs the separate program plus the model weights as two pieces.
- A capable model is several gigabytes to download, and on a laptop without a graphics card it answers slowly.
- The repository has moved:
Mozilla-Ocho/llamafilenow redirects tomozilla-ai/llamafile. Old links still work, so guides written last year point at the old name.
- ollama/ollama
The best-known way to run models locally. It installs a background service and manages a library of models, which is more convenient and less portable than one file.
Track this in Scout - ggml-org/llama.cpp
The engine inside llamafile, used directly. More control, more flags, and you supply the model weights yourself.
Track this in Scout - mudler/LocalAI
It runs models of several kinds behind an API that looks like OpenAI's, so existing code can point at it without changes.
Track this in Scout
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
