Open source AI you can actually ship
If the license says Apache-2.0 or MIT, you can build a commercial product on it, full stop. Everything else in the open-weight world comes with a clause worth reading: user thresholds on Llama and flagship Qwen, derivative rules on older Gemma, and a network clause on AGPL tools that SaaS teams keep rediscovering the hard way. This page maps the clauses to what they mean for a product.
Updated July 12, 2026
The license map
| License | Can you ship? | The catch |
|---|---|---|
Apache-2.0 Includes a patent grant | Yes, unrestricted | Keep the license and NOTICE file in what you distribute; note your modifications |
MIT No patent grant, unlike Apache | Yes, unrestricted | Keep the copyright notice and license text |
Llama Community License Not open source by the OSI definition | Yes, below 700M monthly active users | AUP flows down; derivatives inherit; outputs can't train competing LLMs |
Qwen custom license License varies per model size | Yes, below 100M MAU (flagship sizes) | “Built with Qwen” attribution; most small and mid sizes are plain Apache-2.0 instead |
Gemma terms (pre-4) Gemma 4 switched to plain Apache-2.0 in 2026 | Yes, with flow-down conditions | Broad derivative definition covers distillation and synthetic data; Google keeps remote-restriction rights |
AGPL-3.0 The trap for SaaS products | Yes, but read the network clause | Modifying it and serving users over a network obliges you to publish your modifications |
The user-threshold clauses, in practice
Llama's 700 million monthly-active-user threshold was written to stop a handful of Meta's competitors, and it works exactly that way: the measurement is taken in the calendar month before the model's release, so a startup that grows past the line later keeps its license. Qwen's flagship models carry a similar clause at 100 million MAU, an order of magnitude lower and correspondingly easier to eventually meet. For almost every company reading this page, neither threshold will ever trigger. The real cost is different: these licenses fail the OSI definition, so any enterprise with an open-source-only procurement policy has to route them through legal review instead of waving them through, and many teams simply whitelist unmodified Apache-2.0 and MIT models to skip that conversation entirely.
The clauses that do bite are the quieter ones. Llama's license prohibits using outputs to train competing models, incorporates an acceptable-use policy you must pass on to your users, and has geo-restricted some multimodal releases in the EU. Older Gemma terms define derivatives broadly enough to cover models trained on Gemma-generated synthetic data, with remote-restriction rights on top; Google dropped all of that when Gemma 4 moved to plain Apache-2.0 in 2026, which says something about where the ecosystem is heading.
Derivatives are where teams slip
A fine-tune, a distillation, or in some licenses even a model trained on another model's outputs inherits the parent's terms. The practical consequence: the license of the checkpoint you downloaded is not always the license of the lineage you are in, and the model card on Hugging Face is the only place to confirm it. DeepSeek's R1 distills illustrate the pleasant direction, since the R1 line is MIT and the Qwen-based distills sit under Apache-2.0. The DeepSeek local guide and Qwen local guide note the license per size for exactly this reason.
Tools have licenses too
Model licenses get the attention, but the inference stack around them carries its own terms, and the registry's license data makes the split visible. Three examples spanning the range:
Apache-2.0. The serving engine under many commercial inference stacks, with the patent grant that makes legal teams relax.
MIT. About as unencumbered as software gets; the models it serves carry their own licenses, which is where the review time actually goes.
AGPL-3.0. Fine to use as shipped; modify it and expose it as a service and the network clause obliges you to publish those changes.
The best-known AGPL cautionary tale is Ultralytics YOLO: commercial users who modified it for hosted products faced the choice of publishing their changes or buying the commercial license, which is the dual-licensing business model working as designed.
The EU AI Act deadline is next month
The Act entered into force in August 2024, and its main obligations, including the high-risk rules, apply from August 2, 2026. Open source gets a narrower exemption than most teams assume. AI systems released under free licenses are outside the Act only until they are placed on the market in a high-risk context or interact with users, at which point the transparency rules apply like everyone else's. Open-weight general-purpose models with publicly available weights, architecture, and usage information are excused from some documentation and info-sharing duties, but still owe a copyright policy and a public summary of training data. Models above the systemic-risk compute threshold of 10²⁵ FLOPs get no exemption at all.
The mistake to avoid is assuming the exemption covers your product because the model underneath is open. Deployer obligations attach to the use, not the license: a hiring tool built on an Apache-2.0 model is still a high-risk system.
Questions people actually ask
Is Llama open source?
Not by the OSI or FSF definitions, and both organizations have said so. It is a source-available community license that permits commercial use for almost everyone: the 700 million MAU threshold is measured in the calendar month preceding the model's release, and if you were under it then, your license holds even as you grow. What trips teams up in practice is not the threshold but the flow-down obligations: the acceptable-use policy binds your users, derivatives inherit the license, and the notice requirements travel with the weights.
Can I use an AGPL tool in a commercial product?
Yes, and running it unmodified as a separate service is generally fine. The obligation triggers when you modify the software and let users interact with it over a network: then your modifications must be published under the AGPL. That is why hosted products tend to avoid building on AGPL components they intend to fork, and why Ultralytics YOLO's AGPL license pushed so many commercial users to buy its commercial license instead.
What attribution do I actually owe?
For Apache-2.0: the license text, the NOTICE file if one exists, and a note of your modifications, in whatever you distribute. For MIT: the copyright line and license text. Custom model licenses add their own: Llama requires its copyright notice and license to accompany distributions, and Qwen's custom-licensed sizes ask for a “Built with Qwen” style credit in your documentation.
Does the EU AI Act change any of this?
It adds obligations on top of the license rather than changing it. The Act's main obligations apply from August 2, 2026. Open-source AI systems get a narrow exemption unless they are high-risk, prohibited, or interact with users; open general-purpose models with public weights are exempt from some documentation duties but still owe a training-data summary and a copyright policy, and models past the systemic-risk compute threshold get no exemption at all.
Sources
This is a map, not legal advice: for a product decision, read the exact license file of the exact model version and involve counsel.
The Awesome Open Source AI registry syncs each project's SPDX license from GitHub twice a day, so the license column you see on any project is current rather than copy-pasted from a launch announcement.
