llamafile: a whole model in one runnable file

What is llamafile?

llamafile collapses a language model and the software that runs it into one executable file. Built on llama.cpp and Cosmopolitan Libc, a single file runs across most operating systems and CPU architectures without being installed or even unpacked: you download it, make it executable and run it. It is maintained by Mozilla.ai out of a Mozilla Builders initiative.

At a glance

Product identity

Pricing

Nothing. Apache 2.0, with the llama.cpp and whisper.cpp modifications deliberately kept MIT licensed so they can be contributed back upstream. There is no paid edition, no subscription and no account. Model files are downloaded separately or come packaged inside the executable.

The entire project is free. There is no tier structure and nothing held back. What you supply is a machine, and llamafile is unusually forgiving about which one: it runs on CPU, so a GPU is optional rather than assumed.

Refunds: Nothing is sold, so no refund terms exist.

Platforms

Most operating systems and CPU architectures from the same file, including macOS, Linux, BSD and Windows. Windows imposes a 4GB executable size limit, which matters when a model is packaged inside the file, so larger models are loaded alongside it instead.

• Windows
• macOS
• Linux

Feature matrix

Absence of evidence is not recorded as absence of a feature. Anything that needs an account to confirm is marked not tested rather than "no".

Privacy and data

It cannot do anything with your data, which is the strongest version of this answer. The model runs inside the executable on your own machine, nothing is transmitted, there is no account and there is no vendor. Retention is whatever you choose to save to your own disk. The only caution is about provenance rather than privacy: a llamafile is an executable, and one downloaded from an unknown source deserves the same suspicion as any other program from an unknown source.

Privacy disclosure

What the vendor lets you know before paying

This measures disclosure, not quality. A good product can document nothing, and that is worth knowing separately from whether it is any good.

Content policy and safety

llamafile documents its output filtering asno documented output filter, and its age policy as none stated.

llamafile does not offer clothing removal from photographs of real people. ClickTug does not list tools that do. Its stated position on the likeness of real people is that documented consent is required.

What the vendor prohibits

• Nothing is enforced by the project, which is packaging rather than a model
• What a file refuses is a property of the model weights packaged inside it

General-purpose infrastructure rather than an adult product. It is packaging technology: behaviour belongs entirely to the model someone chose to package, and a llamafile downloaded from a third party is only as trustworthy as its source.

llamafile runs on your own hardware, so it receives none of what you generate. ClickTug does not require a minor safety policy or an acceptable use policy from a tool that never receives user content, because there would be nothing for such a policy to govern. Any optional hosted tier is a separate service under its own terms.

ClickTug content policy

ClickTug verdict

The most portable local AI in this directory and the least like an application. llamafile packages a model and its runtime into a single executable that runs across most operating systems and architectures with no installation, free under Apache 2.0 and maintained by Mozilla.ai. Nothing is transmitted, there is no account and there is no vendor, so the privacy answer is architectural. Against that: a minimal interface, no roleplay or character tooling, a 4GB executable limit on Windows, and the ordinary caution that an executable from an unknown source deserves suspicion.

Hands on

How llamafile compares

Every figure on the right is computed across the 37 tools ClickTug currently tracks, at the moment this page was built. They are facts about this catalogue rather than about the market, and they move as it grows.

Quick answers about llamafile

Generated from the same records as the tables above, so an answer here cannot contradict a figure there. Questions only appear when the answer says something specific to this product.

Is llamafile free?

There is a free tier. The whole project. No paid tier, no trial, no account and no card at any point.

How much does llamafile cost?

No price is published. Free and open source under Apache 2.0, with the llama.cpp and whisper.cpp modifications kept MIT licensed so they remain upstreamable. No paid edition and no account.

Does llamafile require a credit card to start?

No card is required to reach the free layer. One is needed for any paid plan.

Does llamafile offer refunds?

Nothing is sold, so no refund terms exist.

Does llamafile have an app?

No. It runs in a web browser only, which is common for adult products because both app stores restrict them.

Can llamafile run locally?

Yes. It runs models on your own hardware, so prompts and output never leave the machine and the retention question does not arise.

Does llamafile train on my conversations?

The vendor states it does not use user content to train models. That is a written commitment, and rarer in this category than it should be.

Can I delete my llamafile account?

Yes, through a control in the account rather than by asking support.

Can I use llamafile without an account?

Yes. The service can be reached without registering, which almost nothing else in this directory permits.

Is llamafile actually uncensored?

ClickTug records no documented output filter, which is a different claim from having no rules: its terms still prohibit a specific list of conduct, set out above.

What company owns llamafile, and where?

Mozilla.ai, from a Mozilla Builders initiative, Not stated in the repository. Publishing an identifiable counterparty is worth noting: several tools in this directory name no company at all.