"Open source model" is doing a lot of dishonest work in 2026. Half the models wearing the label ship with licenses that restrict who can use them, at what scale, for what purposes — terms that would get any normal software license laughed out of the OSI. If you're building a product, AI model licenses for commercial use are not a formality to skim on download day. They're a dependency with legal teeth, and the time to read them is before your product depends on the weights.
This is a builder's map of the license landscape — what the major license families actually permit, the clauses that bite, who owns model outputs, and a pre-ship checklist. One honest disclaimer up front: this is orientation, not legal advice; for real exposure, pay a real lawyer.
Genuinely permissive (Apache 2.0, MIT). Use commercially, modify, fine-tune, redistribute, embed in products — no user caps, no field-of-use restrictions, no phone-home. Apache 2.0 adds an explicit patent grant, which enterprises care about. Models in this family (large parts of the Qwen line, Mistral's Apache releases, and many others) are the boring, safe choice — and boring is exactly what you want in a dependency.
Community licenses with tripwires (the Llama pattern). Free for commercial use until a threshold — Llama's famous version: if your products exceeded 700 million monthly active users when the model released, you need a separate license. There's also an acceptable-use policy, branding/attribution requirements, and restrictions on using outputs to improve competing models. For most companies the MAU tripwire never fires — but "free unless you become huge" is a term sheet, not open source, and your acquirer's diligence team will read it that way.
Restricted / research-only (non-commercial clauses, RAIL variants, no-derivatives terms). Some model families permit research but bar commercial deployment outright, or attach responsible-AI use restrictions with revocation language, or allow use but forbid distilling outputs into your own models. "The weights are on Hugging Face" does not mean "you may ship this." Downloadable and licensed are different words.
Read past the headline grant and hunt for these: output restrictions (may you train on generated outputs? several licenses say no — this quietly blocks synthetic-data pipelines); redistribution terms (may you ship the weights inside your installer, or must users fetch them from the source? — this decides your offline story); naming requirements (some licenses require derivative models to carry the family name); AUP-by-reference (an acceptable-use policy hosted on a webpage the licensor can edit after you ship); and revocability — permissive grants are effectively irrevocable, while some community licenses reserve termination rights. A license that can be edited or revoked under you is a supplier risk, and you should price it like one.
Two separate questions teams merge into one confusion. Contractually: most licenses and API terms assign output rights to you, sometimes with the no-training-competitors caveat above. Copyright-wise: purely machine-generated content has no author, and authorship is what copyright attaches to — the US Copyright Office has been consistent that AI output without meaningful human contribution isn't registrable. Practical translation: your competitors can't be sued for copying your AI-generated marketing copy under copyright, so your moat is what you build around outputs, not the outputs themselves. Human editing, selection, and arrangement restore protectability in proportion to the humanity involved.
For every model in your stack — including the embedding model everyone forgets, the vision encoder inside your OCR feature, and that quantized variant you grabbed from a random repo (quantizations inherit the original license, whatever the uploader claims): (1) identify the exact license text and version for the exact checkpoint; (2) check commercial permission, thresholds, and field-of-use terms against your actual product; (3) check redistribution rights against your deployment model; (4) check output/distillation clauses against your data pipeline; (5) record it all in an inventory the way you already track npm packages. Run the fast pass with our free AI model license checker, and if the self-hosting decision hinges on economics too, put numbers to it with the self-host calculator. For the strategy layer — which models make business sense to run in-house at all — see local LLMs for business.
A model behind an API can change terms, price, or behavior under you — the license is whatever the dashboard says this quarter. Apache/MIT weights on your own disk are the opposite: an irrevocable capability you own forever, immune to repricing and deprecation notices. That's the licensing philosophy inside QADIR OS — a sovereign, local-first agentic OS built to run permissively-licensed local brains first and treat metered cloud models as optional escalation. We read the licenses so the stack stays shippable: your product, your machine, your rules — enforceable ones.
Ship on rights you actually have. Check any model in seconds with the free model license checker, or join QADIR OS early access — an agentic OS built license-clean from the weights up. No card required.