The model you run,
on terms you’ve read
We build features and products on Llama models, run on your servers, in your cloud or with a host you choose to run them for you, compressed or tuned only as far as your tests allow, and shipped with the line and the files the license asks for, so every term is read before you launch.
Thanks, we’ll be in touch soon.
{{ heroStatus }}
You’re in good company
A few of the companies we’ve worked with.
When the license came last
Many products on Llama models start as a quick download: the model’s files, the part you download and run, put on a spare machine. Someone tunes the model on the team’s own examples, and it ships.
Nobody reads the license that comes with the release, the generation of Llama models those files belong to. So the product shows no “Built with Llama” line anywhere the license names. A tuned version goes to a customer under a name of the team’s own, and the files go out without the license or its notice, the attribution line it asks for in a file named Notice.
Then a customer’s legal team asks which license applies, and nobody can say. Or the team wants a newer model, and its terms are different.
The download proved the model could do the job.
A product needs more: the license read for the way it’s used, and the release, the place it runs, how far its files are compressed to fit smaller machines and any tuning, each picked for a reason.
What we build
Llama models run where your data may go, with the license read for your product, the files kept and every change tested, from one feature to a whole product.
The license, read for your use
What your product must show, hand over and name under the license for its release, read with your legal team before anything launches.
A home for the model
Your servers, your cloud account or a host, a company that runs the model for you, chosen by where your data may go and who will look after it.
Fitted to your machines
The model compressed, its files made smaller to fit the hardware you have, only as far as your tests show its answers still hold.
On phones and laptops
The smaller Llama models run on the device itself, where it can hold them, for features that work without a round trip to a server.
Tuned for your task
A light layer trained on your own examples and added to the model your tests passed, only where tests call for it, and named as the license asks before it’s offered to anyone.
Your own copy of the files
The exact files your tests passed, kept in storage you control with the license and its notice, so the version you tested can always be brought back.
Tests that travel with the model
A test set, your examples and the answers you expect, run on every new release, compression or host before it goes live.
Uses read against the policy
What your feature does, checked against the Acceptable Use Policy, the list that comes with the license of what the models may not be used for, before launch and as the feature grows.
Looked after once it’s live
The servers, the model files and the replies people flag, reviewed on a schedule your team sets.
- Assistants that take actions in your systems?AI Agents
- Answers for customers and staff, from your own documents?AI Chatbots & Assistants
- Your product inside the assistants people use?MCP & Agent Integration
- Models trained on your own data, with their pipelines and monitoring?Data Platforms & MLOps
What the license asks of your product
A product built on a Llama model takes on up to three asks, each set off by something it does: making a product or service with the model in it available to others, handing the model’s files to anyone, and offering a model built from it, or trained on its replies. A product can take on all three. And different releases’ licenses ask different things: an older one asked for a line and a name of its own, and ruled out using its replies to improve other language models.
So we read the license for the release you run with your legal team, place the line where your users can find it, pack the files with the license and the notice, name any model you offer as the license asks, and write down what was done.
Where each ask lands
An example-
Your product’s about page
Asked for when a product or service with a Llama model in it is made available: “Built with Llama”, shown prominently on a related website, user interface, blog post, about page or product documentation.
-
Files you hand over
Whoever gives the model’s files to anyone, a customer’s servers or a device included, gives the license with them, and the notice in a file named Notice, its words taken from the license.
-
A model you offer
A model tuned on a Llama model, or trained on the replies of one, that’s shared or offered to others carries a name beginning with “Llama”. That includes a model trained only on those replies.
Each release has its own license, so we read the one for the release you run.
Every use follows the Acceptable Use Policy that comes with the license.
A company whose products had more monthly users than the license’s threshold, in the month before a release came out, asks Meta for a license first.
For models that read images as well as text, the policy for the newest release doesn’t grant the license’s rights to people or companies based in the European Union. The end users of their products aren’t affected.
Each release comes with its own terms
Llama models come as releases. The files you download stay the same until you change them, so the model in your product is the one your tests passed.
Each release has its own license, and newer releases have asked for different things, so a move is read in its license as well as run on your test set.
Meta also offers newer models, under another name and other terms. Weighing one is a model change like any other: its terms read first, then tried on the copy.
A product on an older release has a clear next step: the license it runs under read, a newer model weighed on its terms and your tests, and a move made only when both hold.
Start with the model you run today
The review covers the Llama releases in use, and the license each comes under; where each one runs, and who looks after it; what the product shows, hands over and names, set against the license; any tuned models, and what they were trained on, other models’ replies included; the files kept, and where; the compression settings and the hardware; the test set, or the cases that could become one; the feature’s uses, set against the Acceptable Use Policy; and where cost can come down.
Then the first changes, in the order we’d make them, starting from a test set.
What you receive is a written review and a proposal for that first step.
Questions worth asking
Does what we send reach the model’s maker?
The model’s files run where you put them. On your own servers or a device, what reaches anyone else depends on the software around the model, which we check before launch. Through a host, the host’s terms decide what’s kept, and we read them with you.
Will a compressed model answer as well?
Sometimes. Making a model’s files smaller to fit your machines can change its answers, so we compress it only as far as your test set allows, and keep the full version to compare against.
Can we use its replies to train another model?
The license for the newest release doesn’t rule it out, and any model you share or offer that way carries a name beginning with “Llama”. An older release’s license ruled it out for other language models. We read your plan against the license for the release you use, with your legal team.
Will the next release come under the same terms?
Not necessarily. Each release has its own license, and newer Meta models come under another name and other terms. We read the terms of any model you’re weighing before you plan a move.
Can a company based in the European Union use these models?
It depends on the models. For models that read images as well as text, the Acceptable Use Policy for the newest release doesn’t grant the license’s rights to a person living in the European Union, or a company whose principal place of business is there. End users of a product that includes those models aren’t affected. We read the policy for the models you’d use with your legal team.
What does it cost to work with you?
It depends on what you’re building, where the model runs and the first step you pick. Before any work starts, you get a written proposal with the scope, the price or rate, and how you pay.
Can we hire developers for our own team?
Yes, through a dedicated team: Webair developers who stay with your product over time and work beside your own people.
Llama and Meta are trademarks of Meta Platforms, Inc. Webair is independent of Meta, and isn’t sponsored by, endorsed by or affiliated with it.
Let’s build it better
Tell us the feature you have in mind and the model it runs on today, and we’ll suggest where to begin, in a short call.
Thanks, we’ll be in touch soon.
{{ ctaStatus }}