What is the difference between permissive and restricted model licenses?

Even when a model’s weights are public, a license still governs what you may do with them. Licenses split into two broad camps, and the gap between them matters for anyone building a product.

Permissive licenses ask for very little. The MIT License lets you use, copy, modify, distribute, and sell the work, including inside closed commercial software, with a single condition: keep the copyright and license notice in the copies you ship.^[2] Apache 2.0 is similar but adds an explicit patent grant. Each contributor grants a “perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable” patent license, so you are protected from later patent claims over the code they contributed.^[1] Apache 2.0 also asks you to note any changes you made and to keep a NOTICE file if one exists.^[1]

Restricted licenses look open because the weights are downloadable, but they add custom rules. Meta’s Llama 3.1 Community License is the common example. If your product exceeds 700 million monthly active users you must request a separate license from Meta, granted at its “sole discretion.”^[3] If you distribute Llama or something built on it, you must “prominently display ‘Built with Llama,’” and any model you train from it must start its name with “Llama.”^[3] Use is also bound to an Acceptable Use Policy.^[3]

These extra terms are why many people argue these licenses “cannot reasonably be called ‘open source.’”^[4] The restrictions are not unusual or hidden, but they are real legal hurdles, and they are absent from standard licenses like Apache 2.0 and MIT.^[4]

Llama is not the only case, and the restrictions are not all the same shape. Google’s Gemma license reserves Google the right to “restrict (remotely or otherwise)” any use that violates its prohibited-use policy — a live kill switch a permissive license simply doesn’t have — and extends that same restriction to any model trained on synthetic data Gemma itself generated, which developers have flagged as making commercial use riskier than it first looks.^[4] The specific catch differs by vendor: a user-count cap here, a remote kill switch there, a ban elsewhere on using a model’s outputs to improve anything but that vendor’s own line. The pattern underneath is the same one worth checking every time — “the weights are downloadable” and “the license leaves you alone” are two different claims.