Llama

What Is Llama?

Llama is Meta's family of large language models, published as downloadable weights rather than sold as a service. You can run it on your own hardware, fine-tune it on your own data, and deploy it without telling anyone. More than any other single thing, Llama is why running serious AI locally went from a research curiosity to something a person can do on a laptop.

It is also the model most other things are built on. A very large share of the fine-tuned, quantised, specialised models circulating in the open ecosystem started life here, and much of the tooling for local inference exists because Llama gave it something worth running.

There is one thing about it that consistently surprises people, and it is not technical. Llama is free, but it is not open source. That distinction has real consequences, and it is worth understanding before you build a company on top of it.

Where the Name Went

Worth knowing before you go looking: llama.com now redirects to Meta's developer site, and Meta's newest models carry a different name entirely. The current line-up is branded Muse — a coding-focused agentic model, a companion multi-agent terminal tool, and a compact open model built to run local agents on a single consumer GPU. There is a self-serve Meta Model API alongside them, compatible with OpenAI-style clients.

Llama itself has not gone anywhere. Llama 3 and Llama 4 remain documented and downloadable, and the enormous installed base of deployments and derivatives is unaffected. But if you are choosing a Meta model today rather than maintaining one, the newest work is happening under the Muse name, and the open-weights tradition continues there rather than under the old one.

What Llama 4 Actually Offers

The current Llama generation is natively multimodal — text and vision trained together rather than bolted onto each other — and comes in two shapes.

Scout is the efficiency play: multimodal, designed to run on a single high-end datacentre GPU, and carrying a context window measured in millions of tokens. For long-document work, that combination is the reason to look here rather than at a hosted API.

Maverick is the capability play: multimodal image and text understanding aimed at high throughput and low cost, with Meta's own estimate placing inference at around nineteen cents per million tokens under distributed serving. That figure assumes you are running it properly at scale, which is a meaningful assumption.

Both are strong on document understanding and image reasoning and competitive rather than dominant on coding. As always, benchmark tables from a vendor tell you what the vendor wants measured — test on your own material before committing.

The Licence: Read This Part

Llama ships under Meta's own Community Licence, not under Apache, MIT or any other standard open-source licence. It is permissive enough for almost everybody and it carries specific conditions that almost nobody reads.

• If your product exceeds 700 million monthly active users at the time of release, you must request a licence from Meta, which Meta may grant at its sole discretion. This clause exists to exclude Meta's largest competitors, and it will never affect you — but it is the reason the licence is not open source.
• You must display «Built with Llama» prominently on any product or interface that uses it.
• Any derivative model you distribute must have a name beginning with «Llama».
• You must include a copy of the agreement and the copyright notice with anything you distribute.
• Meta's Acceptable Use Policy applies to what you do with it, alongside ordinary legal compliance.
• You own the derivative works you create; Meta retains ownership of the original materials.

None of this stops a normal business from using Llama commercially. But the attribution and naming clauses are contractual obligations rather than courtesies, and a company that quietly ships a fine-tuned Llama under its own brand with no attribution is in breach. If that describes your plan, either read the licence properly or choose a model under a genuine open-source licence.

What It Costs

The weights are free. Everything else is not.

Running a large model yourself means GPUs, either bought or rented, plus the engineering time to serve it reliably — batching, quantisation, monitoring, updates. For a small team the honest comparison is not «free versus twenty dollars a month» but «a salary and a server versus an API bill», and the API usually wins until volume is substantial.

The middle path is what most people actually use: Llama hosted by somebody else. Every major cloud and most inference providers serve it, usually at prices well below the closed frontier models, because they are competing on the same freely available weights. You get the cost advantage without the operations burden, at the price of the data-locality benefit that made you interested in the first place.

Genuinely free is the small end: a compact Llama on a modern laptop, running offline, costing nothing but disk space. For summarising, drafting, classification and private work, that is a real answer rather than a demo.

Who Gets the Most From It

Anyone Whose Data Cannot Leave

Healthcare, legal, defence, finance, and any organisation with a policy rather than a preference. Downloadable weights turn «we are not allowed to use AI» into an infrastructure question, and that is the single most valuable thing Llama does.

Teams Fine-Tuning on Their Own Material

A model shaped by your documents, your tone and your domain, running where you choose, owned by you. Prompting a closed API cannot reach the same place, and the derivative is yours to keep.

Developers Building Local or Offline Products

Applications that work on a plane, on a factory floor, or in a country with unreliable connectivity. Compact models running on-device make an entire category of product possible that a cloud API simply cannot serve.

People Learning How Models Work

You can inspect it, break it, quantise it, and watch what happens. For anyone who wants to understand these systems rather than only use them, having the weights is not a convenience — it is the whole possibility.

What to Watch Out For

• Free is not open source. The Community Licence carries attribution, naming and user-threshold conditions that bind you contractually.
• llama.com now redirects to Meta's developer site, and Meta's newest models are branded Muse rather than Llama.
• Self-hosting a large model is an engineering project. The licence costs nothing; the GPUs and the person who keeps them running cost a great deal.
• The top closed models still generally lead on the hardest reasoning tasks. Choose Llama for control and cost, not for winning benchmarks.
• Vendor benchmark tables measure what the vendor chose to measure. Test on your own workload.
• «Built with Llama» must appear on your product. Plan for it in your interface rather than discovering it in legal review.
• Hosted Llama gives you the price advantage but not the privacy advantage — your data still goes to somebody else's server.
• Model generations move fast and community fine-tunes move faster. Whatever you deploy today will be superseded; design for replacement.

Frequently Asked Questions

Is Llama free?

The weights are free to download and use commercially under Meta's Community Licence. Running them is not free — you pay in hardware or in hosting — and the licence attaches conditions, so free does not mean unconditional.

Is Llama open source?

No. It ships under a proprietary community licence with attribution requirements, a naming rule for derivatives and a restriction on products above 700 million monthly active users. Widely available, but not open source in the accepted sense of the term.

Can I use it in a commercial product?

Yes, provided you stay under the user threshold, display «Built with Llama», prefix any derivative model's name with «Llama», include the licence and copyright notice with anything you distribute, and follow the acceptable use policy.

What happened to llama.com?

It redirects to Meta's developer site. Llama 3 and Llama 4 are still documented and downloadable there, but Meta's newest models are branded Muse — including a compact open model built for local agents.

What hardware do I need?

It depends entirely on size. A compact quantised model runs on a decent laptop. The larger multimodal models are designed around high-end datacentre GPUs — one for the efficiency-focused variant, more for serving at scale.

How does it compare to running GPT or Claude?

You cannot run those at all — that is the comparison. On capability, the closed frontier models generally lead. On control, privacy, cost at volume and freedom from a vendor's decisions, Llama wins by default because the alternatives do not compete in that category.

The Bottom Line

Llama changed what is possible for anyone who could not or would not send their data to somebody else's server. It remains the default starting point for local and self-hosted AI, and the ecosystem built on top of it is larger than the model itself.

Two things to carry with you. The licence is not open source and its attribution clauses are real obligations, so read it before you ship. And Meta's newest work now appears under the Muse name, so check which family you are actually choosing rather than reaching for the one you remember.