Most people download a model, run it, and never open the licence. That is fine right up until the moment it is not: you build something on a model, it starts working, and only then do you find out what the terms said. This guide is the two-minute routine we run before a download finishes, and what the answers actually mean.

We are hobbyists who run models on one desktop card, not lawyers. Nothing here is legal advice. What it is: a description of what these specific licences say, so you know when it is worth asking somebody who is a lawyer.

The three things people call “open”

These get used interchangeably and they are not the same.

Open source has an actual definition, maintained by the Open Source Initiative, and it includes the right to use the thing for any purpose. Very few large models qualify.

Open weights means the trained model file is published. You can download it, run it offline, fine-tune it, and inspect it. What you may legally do with the output and the model is set by its licence, which may or may not be an open-source one. This is the category most “open” models actually sit in, and it is the one that matters for running things locally.

API-only means you can reach the model through somebody’s service and nowhere else. The weights are not published. You are renting access. A model can be free to use and still be API-only.

A model page can be enthusiastic about “open” in the headline and be API-only in the fine print. Both happened in one week in August 2026, at the same company.

The routine

  1. Find the actual LICENSE file, not the model card and not the announcement post. On Hugging Face it sits in the file list at the repository root.
  2. Read the name and version. “Apache 2.0” and “Qwen Community Licence 1.0” are wildly different documents, and the second one is not the first.
  3. Search the text for a threshold, a number of users, or an amount of revenue. Many custom model licences are permissive below a line and restrictive above it.
  4. Search for the words “service”, “hosted”, “API”, and “assistant”. This is where the clauses live that catch people who build products rather than run things at home.
  5. Ask one question: which side of this am I on? For most people running a model on their own machine, the answer is the permissive side, and you can stop worrying. If you are reselling access or shipping a product, the answer may be different and it is worth an hour of a lawyer’s time before you build.

Worked example: the Qwen Community Licence 1.0

Qwen ships some models under Apache 2.0 and others under their own licence, so the family name tells you nothing. Qwen3.8-Flash-Next is under the Qwen Community Licence 1.0. Here is what that document actually says.

There is a threshold, and it is an attribution requirement. Above 100 million monthly active users or US$20 million in monthly revenue, you have to display the model’s name. That is it at that tier: say what you are using. If you are reading this guide, that clause is not about you.

Two things have no threshold at all. If you run what the licence calls model as a service, defined as giving a third party access to language model inference or fine-tuning, for example through an API or a hosted endpoint, then you need a separate licence from Qwen before any commercial use. The same applies if you build an AI work assistant, which the licence defines as an independent product primarily for AI-assisted coding or office productivity. Single-purpose tools and domain-specific assistants are carved out of that second category.

So the practical read: running it on your own machine, including for your own commercial work, is not what these clauses are aimed at. Standing up an endpoint other people pay to call, or shipping a coding assistant, is.

That is a real restriction and it is worth knowing. It is also not the same as “you cannot use this”, which is how licence changes usually get reported.

Worked example: when the page says no

Wan 3.0 took the number one spot on Artificial Analysis for text-to-image in August 2026. It is API-only. Alibaba’s own page carries a line reading “Open Source: No”. Several third-party sites listed Apache weights for it, and following those claims led back to nothing official.

The same week, the same company published Wan 2.2-Animate-2 at 14B, genuinely open under Apache, on Hugging Face.

Both of those are true at once, and together they show the shape of the current era: the open lane tends to run about one generation behind the API lane. That is not a betrayal, it is the deal on offer, and it is worth knowing you are in it. The failure mode is not a company keeping its best model behind an API. The failure mode is you assuming it did not.

When a third-party page claims weights and the vendor’s own page does not, believe the vendor.

What Apache 2.0 actually gives you

Since it is the benchmark everything else gets compared to: Apache 2.0 lets you use, modify, distribute, and sell, commercially, with no user or revenue threshold. You must keep the licence and notice files and state significant changes you made. It also includes an express patent grant, which is the part that matters to companies and that most people never read.

IBM’s Granite 4.2 shipped under it in August 2026. So did Wan 2.2-Animate-2. When you see Apache 2.0 on a model page, the licence question is essentially answered and you can go back to worrying about VRAM.

The habit worth building

Open the licence file before the download finishes. It takes two minutes, it is usually good news, and the one time it is not, you will have found out before you built something on top of it.

We run all of this locally on one desktop GPU and publish what we find, including the parts we get wrong. Everything here is a description of what these licence documents say, not legal advice.