Open-Weights Model Licenses Explained: Apache, MIT, and Community Terms

Downloading a model takes one command. Working out whether you are allowed to use its output in your business takes considerably longer, and most people skip it.

That skip is usually harmless and occasionally expensive. Model licenses in 2026 range from genuinely permissive to flatly incompatible with commercial work, and the download experience is identical either way. Here is what the main categories actually permit.

Apache 2.0: The Cleanest Option

Apache 2.0 is the license you want to see. It permits commercial use, modification, redistribution, and private use, with two obligations: keep the license and attribution notices, and state significant changes if you redistribute a modified version.

There is no user-count threshold, no field-of-use restriction, and no acceptable-use policy bolted on. It also includes an express patent grant, which is the technical reason lawyers prefer it to MIT for anything substantial.

As of mid-2026, Apache 2.0 covers a large share of the practical open-weights field: OpenAI’s gpt-oss models, Google’s Gemma 4 family, Alibaba’s Qwen3 releases, and most of Mistral’s open lineup.

MIT: Shorter, Almost As Good

The MIT license is about two hundred words and grants essentially unrestricted use with attribution. No patent grant, which matters to some legal departments and not at all to most users.

Several of the large Chinese open-weights releases ship MIT, including models in the DeepSeek and GLM lines. For a company using a model rather than redistributing it, MIT and Apache 2.0 are practically equivalent.

Community Licenses: Usable With Conditions

Meta’s Llama family popularized the bespoke community license, and several labs have copied the pattern. These are not open source by the OSI definition, and the differences are worth knowing:

A user-count ceiling. The Llama community license requires a separate license from Meta if your product exceeds a very large monthly-active-user threshold. Irrelevant for a forty-person consultancy, relevant for a consumer app at scale.

An acceptable-use policy. A separate document listing prohibited uses. It is incorporated by reference, which means it binds you even though it is not in the license file, and it can be updated.

Naming and attribution requirements. Derivatives often must carry the model family name and reproduce license text.

None of this makes community-licensed models unusable. It does mean you cannot treat them as drop-in equivalents to Apache 2.0 when you are shipping software.

Research-Only Licenses: The Real Trap

Some models are published under terms that permit research and evaluation but forbid commercial use. They appear on the same model hubs, download with the same command, and produce output indistinguishable from a permissively licensed model.

This is the failure mode that actually catches teams, because nothing about the technical experience warns you. Wrivio hit it: an earlier build shipped a model under a non-commercial research license, and the fix was replacing it with Apache 2.0 alternatives rather than negotiating terms.

If you are selecting a model for anything that touches revenue, the license check is not paperwork. It is the first step.

Weights You Hold Cannot Be Repriced

One underrated property of open weights: a file on your disk is not subject to next quarter’s terms of service.

API providers change pricing, deprecate models, adjust rate limits, and revise data-handling terms, all legitimately and all on their schedule. A locally stored model under Apache 2.0 keeps working under the license you accepted, on hardware you control, indefinitely.

For a writing tool that people build habits around, that stability is worth as much as the privacy. It is why Wrivio’s Local mode ships weights to your machine rather than proxying a hosted model.

A Short Checklist Before You Adopt A Model

Run these in order. The first two catch nearly everything.

1. Read the license file itself, not the launch blog post.
2. Confirm commercial use is permitted for your actual use case.
3. Check for a separate acceptable-use policy incorporated by reference.
4. If you will redistribute or bundle the model, confirm distribution rights and any naming obligations.
5. Note the license in your own records, with the model version and the date you checked.

That last step sounds bureaucratic until an auditor asks which model version processed client data in March. We wrote about that specific problem in how to document your AI workflow for an auditor.

Writing The Adoption Request

If you need sign-off to use a model at work, the license facts belong in the request, in plain language rather than legalese.

Before:

Hey, can I use a local AI model for rewriting emails? It’s open source so it should be fine.

After:

I would like approval to use a local open-weights model for rewriting internal and client email. The model is licensed Apache 2.0, which permits commercial use with attribution and carries no user-count restriction. It runs entirely on my workstation, so no message text is transmitted to a third party. I can provide the license text and the model version for your records.

The second version answers the two questions Legal will ask anyway. A Wrivio Context for approval requests could say:

Rewrite this as a formal internal approval request. Professional register, complete sentences, no contractions. Lead with the specific ask. Keep every license name, version number, and technical claim exactly as written. Do not add commitments or benefits that are not in the original.

If you want a fuller template, see how to ask your company to approve an AI tool.

Common Questions

Does an Apache 2.0 model mean the training data was licensed?

No, and this is an important gap. The license covers the weights, not the provenance of what trained them. Open weights tells you about your rights; it tells you nothing about the lab’s data sourcing.

Do I own the text a model produces?

Model licenses generally do not claim rights in output, and permissive licenses are explicit about this. Copyright in AI-assisted output is a separate and unsettled question that varies by jurisdiction, and it is worth a conversation with counsel if the stakes are high.

Can a lab revoke a license after I download the weights?

The lab can change the terms for future downloads. Weights you already hold under an accepted grant are generally yours to keep using under those terms, which is one reason to record the license version alongside the model.

Is a community license risky for a small business?

Usually not. The clauses that bite are aimed at very large deployments. Read the acceptable-use policy, confirm your use is not on the prohibited list, and keep a copy of both documents.

Download Wrivio for Windows to run an Apache 2.0 open-weights model locally, with no per-token cost and no terms that change under you.