Written by Jakub Rusinowski · Last updated 2026-07-12 · Hardware figures computed by our VRAM engine
This page is practitioner guidance on deployment mechanics, not legal advice — involve your counsel or DPO for decisions about your specific obligations.
Open-weight model licenses fall into two families: true open-source licenses (Apache 2.0 — Qwen, Mistral Small, GPT-oss, Granite; MIT — DeepSeek, GLM) that permit commercial use with minimal conditions, and vendor community licenses (Meta's Llama license, Google's Gemma terms) that permit most commercial use but attach conditions — acceptable-use policies, attribution, and in Llama's case a threshold clause for very large companies. For most businesses all of these are workable; the difference is how many questions legal has to ask first.
| Standard open source (Apache 2.0, MIT) | Vendor community licenses (Llama, Gemma) | |
|---|---|---|
| Commercial use | Yes, unconditionally | Yes, with conditions |
| Example models | Qwen 3, Mistral Small, GPT-oss, Granite (Apache 2.0); DeepSeek, GLM (MIT) | Llama 3.x / 4 (Llama Community License); Gemma (Gemma Terms of Use) |
| Acceptable-use policy | None baked into the license | Yes — flows down to your deployment |
| Attribution | Keep license/notice files (Apache adds a NOTICE requirement) | Yes — e.g. Llama requires "Built with Llama" display for distributed products and llama-prefixed names for derivative models |
| Patent grant | Apache 2.0: explicit; MIT: no express grant | License-specific terms |
| Special clauses | None | Llama: companies with >700M monthly active users at release date need a separate Meta license |
| Legal review effort | Minutes — these licenses are decades-understood | Hours — someone must actually read the AUP and terms |
The practical summary most legal teams land on: Apache 2.0 and MIT models are approve-once, the same review as any open-source dependency. Community-licensed models are approve-per-use-case, because the acceptable-use policy is a living document you're agreeing to enforce — and vendors can revise it for future model versions (each model release binds you to the license it shipped with; already-downloaded weights don't retroactively change).
1. "Can we use the outputs commercially, and who owns them?" All the mainstream licenses permit commercial use of outputs. Ownership of outputs is a broader unsettled area of law (AI-generated content and copyright), but nothing in these licenses claims your outputs for the vendor. Llama's license notably permits using its outputs to improve *other* models in current versions — a restriction earlier drafts were stricter about; check the version you're actually using.
2. "What must we display or keep?" Apache 2.0: retain license and NOTICE files in anything you redistribute — invisible for internal deployments. Llama: products *distributed* to third parties display "Built with Llama," and fine-tuned derivatives carry "Llama" in their name. A purely internal assistant triggers little of this; an AI feature in your shipped product triggers all of it.
3. "What does the AUP forbid, and can we enforce it on our users?" Community licenses prohibit lists of uses (illegal activity, some regulated domains, deception). For an internal tool with authenticated users this is manageable policy work — your usage policy mirrors the AUP, your gateway logs evidence it. For a customer-facing product, you are effectively flowing Meta's or Google's policy down to your customers; legal will want to see that reflected in your terms of service.
4. "What happens to our fine-tunes?" Under Apache 2.0/MIT: your fine-tuned weights are yours, redistribute or keep private as you wish. Under Llama: derivatives remain governed by the Llama license (naming, AUP, and all), which matters if you ever ship or sell the tuned model. This is also vendor-checklist question #7 — a *platform vendor* claiming ownership of fine-tunes you trained on an open model is a contract problem, not a license inevitability.
5. "Is there any indemnification?" No. No open-weight license — standard or community — indemnifies you for outputs, infringement claims, or anything else; that's a paid-API feature some cloud providers offer. Self-hosting trades that indemnity for control. Most legal teams accept this the way they accept it for every other open-source component in the stack, with the same mitigations: human review where outputs matter, and insurance where it doesn't suffice.
The policy that keeps procurement simple, used implicitly across this hub's recommendations: