Is Llama open source?

Not in the Open Source Initiative sense. Llama is released under a custom community license that permits broad commercial use but adds conditions an OSI-approved license cannot contain — an acceptable use policy restricting certain applications, an attribution and naming requirement for derivative models, and a monthly-active-user threshold above which a separate license from Meta is required. The accurate term is open weights, not open source.

Can I use an open-weight model commercially without paying?

Usually yes, but conditionally. Most open-weight licenses grant a royalty-free commercial right subject to compliance with the license's use restrictions, attribution terms, and any scale trigger. The cost is compliance obligations rather than fees — which is a different risk profile, not an absent one.

Do license terms follow my fine-tuned model?

Generally yes. Open-weight licenses typically define derivative works to include fine-tunes, distillations, and merged models, and require that the same terms be passed downstream when you distribute them. Several also require the base model's name to appear in your derivative's name and require redistribution to carry the original license file and notice.

What is the difference between an open-weight license and Apache 2.0?

Apache 2.0 is a true open-source license: it imposes no field-of-use restrictions, no user thresholds, and no naming requirements beyond notice preservation. Some model families ship under Apache 2.0, which is materially simpler to comply with. Custom community licenses layer use restrictions and conditions on top, which is precisely what disqualifies them from the open-source label.

Does the model license cover the outputs my product generates?

Model licenses generally do not claim ownership of outputs, but several extend the acceptable use policy to how outputs are used, and some restrict using outputs to train competing models. Read the outputs clause separately from the weights clause — they are distinct grants with distinct conditions, and the output-to-train restriction is the one most commonly violated by accident.

Open-Source AI Model Licenses 2026: What 'Open' Actually Obligates You To

"Open Weights" and "Open Source" Are Different Products

An open-source license, by definition, cannot restrict fields of use. It cannot say "not for military applications," cannot condition the grant on how many users you have, and cannot require you to name your product after the licensor. Every one of those conditions appears in at least one widely used model license.

This is not a criticism of the licenses — publishing weights with conditions is a legitimate choice and vastly more permissive than closed APIs. The problem is vocabulary. Engineering teams hear "open source," apply the mental model they use for a Python package, and skip the review step. Legal finds out during a diligence process two years later.

The Four Clause Types That Actually Bite

Attribution and Naming

•Display a 'Built with X' notice on the product surface
•Include the base model name in your derivative's name
•Ship the original license file and notice on redistribution
•Preserve copyright attribution in documentation

Acceptable Use Policy

•Incorporated by reference — and can be updated by licensor
•Typically bans specific harmful application categories
•Often flows down to your end users via your own terms
•Violation can terminate the license entirely

Derivative and Output Restrictions

•Fine-tunes and distillations inherit the license
•Some bar using outputs to train competing models
•Merged models can inherit terms from every parent
•Redistribution must pass the same terms downstream

Scale and Territory Triggers

•User-count thresholds requiring a separate license
•Measured at a fixed date, not continuously
•Some research licenses bar commercial use outright
•Region-specific carve-outs in a few model families

The Same Vendor Ships Multiple License Tiers

A common and expensive mistake: assuming a model family has one license. Several labs publish some models under a permissive open-source license and others — usually the larger or newer ones — under a restrictive research-only license that forbids commercial deployment entirely. The model cards sit next to each other on the same hub page, download through the same code path, and differ only in a license field most fetch scripts never read.

Treat the license as a per-checkpoint attribute, not a per-vendor one. A team that cleared "Vendor X models" in a legal review and later swapped in a newer checkpoint from the same vendor has not actually cleared anything.

Where This Surfaces: Acquisition and Enterprise Diligence

Model license compliance rarely produces a lawsuit. It produces a diligence finding. Enterprise security questionnaires now routinely ask which models are deployed and under what terms, and acquirers' IP counsel will ask whether the fine-tuned weights that constitute a large share of the company's technical value can be transferred and on what conditions. Discovering mid-transaction that your core model inherits a naming requirement or a research-only restriction is a valuation conversation, not a compliance memo.

Compliance Checklist

Inventory and Classification

☐List every model checkpoint in production, by exact version
☐Record the license per checkpoint, not per vendor
☐Flag research-only and non-commercial licenses immediately
☐Trace merged models back to every parent's terms

Product and Process Controls

☐Place required attribution notices in the product UI
☐Flow acceptable-use restrictions into your own customer terms
☐Add a license check to the model-onboarding pipeline
☐Re-review when upgrading to a new checkpoint

Frequently Asked Questions

We only call the model through a hosted inference provider. Are we still bound?

Usually you are bound by the provider's terms rather than the model license directly, but the underlying use restrictions typically flow through. Check whether your provider's terms incorporate the model's acceptable use policy — most do — and whether they pass the attribution obligation to you or absorb it themselves.

Does the acceptable use policy apply to our end users or just to us?

Both, in practice. These licenses commonly require you to pass the restrictions to anyone you make the model or its outputs available to. That means your own terms of service need corresponding language, and enforcement is your responsibility, not the licensor's.

If a licensor updates its acceptable use policy, are we bound by the new version?

Often yes — many licenses incorporate the policy by reference to a live URL rather than pinning a version. That is an unusual and under-appreciated term: your obligations can change without you signing anything. Snapshot the policy at adoption and monitor it, the same way you would monitor a critical dependency.

Are the naming requirements really enforced?

Enforcement to date has been light, but the risk is not the licensor suing you — it is that non-compliance is trivially verifiable from the outside. A model card, a product page, or a Hugging Face repo either carries the required notice or it does not. That makes it an easy finding for anyone reviewing you.

Is there a genuinely simple option?

Yes — model families released under Apache 2.0 or MIT. They carry ordinary open-source obligations: preserve notice, and that is largely it. If license overhead is a real constraint for your team, weighting checkpoint selection toward permissively licensed families is a legitimate architectural decision, not a legal-department preference.

Read the License Before the Benchmark

Model selection conversations start with evaluation scores and inference cost, and the license shows up — if at all — after the integration is built. Reversing that order costs nothing at the start and can save a rewrite later, because license terms are the one model attribute you cannot optimize your way around.

Add a license field to whatever document your team uses to compare candidate models. Two lines: the license name, and whether anything in it touches your product surface.