The Model Is Free. What Does the License Cost?

The most consequential document in your AI stack right now might be a license nobody in your company has read. Every open-weight model ships under one, and the range is wider than most leaders assume: some licenses genuinely let you do anything, while others carry revenue thresholds, attribution mandates, naming rules, and termination clauses that can require you to delete the model your workflows run on. We call the boundary those terms draw the license line, and it's the second line every open-model decision has to cross. The weights tell you what you hold. The license tells you what you're allowed to do with it, and holding is the easy half.

What is the license line?

It's the boundary between possessing a model and being permitted to use it the way your business intends.

We've written about the weight line, the question of whether you hold a model's weights or rent access to someone else's. The license line sits directly behind it. A model can clear the weight line completely, sitting on your own hardware, fully under your operational control, and still carry terms that constrain what you build, what you name it, what you display in your product, and what happens if you grow. This is also a different exposure from the availability clause, which covers hosted access ending on a timeline you don't choose. The license line governs the copy you already downloaded. Nobody has to switch anything off for it to bind you. You agreed to it the moment you clicked accept, and in most companies, the person who clicked was a developer moving fast on a Tuesday.

What does a permissive license actually permit?

Everything that matters, and the entire text fits on a single page.

The MIT license, which is how Z.ai ships GLM-5.2, grants the right to use, copy, modify, merge, sell, and sublicense the software, free of charge, on one condition: keep the copyright notice attached. That's the whole deal. No revenue triggers, no attribution requirements in your product, no naming rules, no separate agreements waiting at any scale. Apache 2.0, the other workhorse of this tier, adds an explicit patent grant on top. When leaders picture "free and open," this tier is the thing they're picturing, and the picture is accurate here. Your counsel can read the full terms over coffee, and what they'll tell you is that the license imposes almost nothing. For production systems, that near-nothing is worth real money, because every condition a license doesn't contain is a condition that can never surprise you later.

Where do "open" licenses attach strings?

At the point where you succeed. The conditions in this tier don't price your present. They price your growth.

Meta's Llama models ship under a community license that reads like a formality and isn't. It requires "Built with Llama" displayed on your website or product documentation, requires any model you fine-tune from it to carry "Llama" at the start of its name, incorporates an acceptable use policy by reference, and, above 700 million monthly active users, converts your free license into one Meta may grant at its sole discretion. Moonshot's Kimi K3, the biggest open-weight release of the summer, arrived under a custom license with the same shape: free for nearly everyone, with a separate agreement required once a model-as-a-service business crosses 20 million dollars in revenue, and interface attribution required past 100 million monthly users. Most companies will never touch those thresholds, which is exactly why the clauses get skimmed. Two kinds of companies should read them anyway: the ones planning to grow, and the ones planning to be acquired, because acquisition diligence reads every license you didn't, and a conditional license on a load-bearing model is the kind of finding that reprices a deal.

Which clause deserves the most attention?

Termination, because it's an off switch you signed yourself.

Buried in Llama's community license is a clause that terminates your rights the day you file an intellectual property claim against Meta, at which point you're obligated to delete the model and stop using it. Read that again from an operator's chair: a legal dispute in one part of your business can create a contractual duty to remove infrastructure from another. Use policies incorporated by reference deserve the same scrutiny, because they can be updated after you deploy, and your compliance obligation moves with them. None of this makes conditional licenses unusable. Plenty of excellent systems run on Llama models inside the terms. It makes them contracts, and contracts get read, tracked, and weighed before anything load-bearing gets built on top of them.

Does the license on your tools cover the model?

No, and this is the gap that catches technical teams.

The runtime and the model are licensed separately. Ollama, the most popular way to run local models, is itself MIT-licensed, and that clears nothing about the model you load into it. A permissively licensed tool serving a conditionally licensed model leaves every condition intact. The same is true across the stack: your orchestration framework, your serving layer, and each model you deploy all carry their own terms, and clearing one layer clears only that layer. When your team says "it's all open source," the correct follow-up is "under which licenses," asked once per component. It's a five-minute question that most companies have never asked even once.

How do you keep a license from becoming a liability?

Treat model licenses as what they are: supplier contracts. Then run them through the process you already have for supplier contracts.

The working approach we recommend has three parts, and none of them require a legal department on retainer. Inventory first: list every model in production or in pilot, with its exact license attached, including the ones developers pulled down to experiment with, because experiments have a way of shipping. Then sort the list into permissive and conditional, and for everything conditional, put the thresholds next to your growth plan and any exit scenario, and have someone with contract literacy read the termination language before the model touches production, in the same gate where you already run security review. Last, make the license a tiebreaker: when two models clear the same job, and at most workloads they now do, the one under MIT or Apache terms is carrying less future risk at the same capability, and that difference is worth choosing on purpose. The companies that get burned here won't be the ones that used conditional models. They'll be the ones that found out the conditions existed from an acquirer's counsel.

If you want your model stack inventoried, sorted, and built on licenses that won't surprise you later, learn about our AI Blueprint approach or reach us at contact@theyor.com.