Open-source AI has a licensing problem. The term “open source” gets applied to models with wildly different legal terms, from truly permissive Apache 2.0 to restrictive custom licenses that barely qualify as open. Choosing a model without understanding its license is a legal risk that too many teams take.<\/p>\n
Apache 2.0 is what most people mean when they say “open source.” It grants permission to use, modify, distribute, and sell the model and anything built with it. No revenue caps, no usage restrictions, no mandatory attribution in the product (though you must include the license in source distributions).<\/p>\n
Models under Apache 2.0 include Mistral’s releases, many Qwen variants, and several smaller but capable models. For commercial use, Apache 2.0 is the safest choice. Your legal team will sign off immediately because the terms are well-understood and battle-tested across decades of software licensing.<\/p>\n
The key clause: Apache 2.0 includes a patent grant<\/strong>. Contributors implicitly license any patents that cover their contributions. This matters for organizations concerned about patent litigation \u2014 it provides a layer of legal protection that more permissive licenses like MIT don’t include.<\/p>\n Meta’s Llama models use a custom license that looks open but has important restrictions. The headline limitation: if your product or service has more than 700 million monthly active users<\/strong>, you need a separate commercial license from Meta.<\/p>\n For 99.9% of companies, this threshold is irrelevant. But for large platforms \u2014 social networks, search engines, major cloud providers \u2014 it creates a dependency on Meta’s commercial terms. The license also prohibits using Llama outputs to train competing models, which limits certain research use cases.<\/p>\n The Llama license permits commercial use, modification, and redistribution below the user threshold. Fine-tuned models and merged models inherit the license. If you merge a Llama model with an Apache 2.0 model, the result carries Llama’s restrictions \u2014 the more restrictive license wins.<\/p>\n Legal teams sometimes struggle with the Llama license because it’s bespoke. Unlike Apache 2.0 or MIT, there’s limited legal precedent for interpreting its terms. Ambiguous edge cases \u2014 like what counts as a “monthly active user” for an API service \u2014 create uncertainty that risk-averse organizations dislike.<\/p>\n Alibaba’s Qwen models have shifted licensing over time. Earlier versions used custom licenses; recent releases moved toward Apache 2.0, which is a significant improvement for commercial adoption. However, some Qwen variants still carry restrictions worth checking.<\/p>\n The geopolitical dimension matters. Some organizations have policies against using models developed by Chinese companies, regardless of the license terms. This isn’t a legal issue but a compliance and risk management consideration that affects model selection in certain industries, particularly defense, government, and some financial services.<\/p>\n From a pure licensing perspective, Qwen models under Apache 2.0 are as permissive as any other Apache-licensed model. The weights are the weights; the license is the license. Technical evaluation should be separate from geopolitical considerations.<\/p>\n Gemma’s license<\/strong> (Google) is relatively permissive but includes a prohibition on using the model to generate training data for other models and restrictions on deploying in certain sensitive applications. It’s more permissive than Llama’s license but less clean than Apache 2.0.<\/p>\n DeepSeek’s license<\/strong> permits research and commercial use but includes restrictions on using outputs for model training. The terms are somewhat ambiguous on derivative works, which creates questions for fine-tuning use cases.<\/p>\n Creative Commons licenses<\/strong> occasionally appear on datasets but aren’t well-suited for software or model weights. CC-BY-NC (non-commercial) in particular causes problems \u2014 models trained on NC-licensed data inherit the non-commercial restriction, which can spread through the ecosystem like a virus.<\/p>\n The Open Source Initiative (OSI) has been working on defining what “open source” means for AI models. Their position is clear: models with usage restrictions, revenue caps, or field-of-use limitations are not open source<\/strong> by the traditional definition, regardless of how they’re marketed.<\/p>\n This creates a vocabulary problem. The community uses “open source” loosely to mean “weights are downloadable,” while the OSI definition requires freedoms that many popular models don’t grant. Terms like “open weights,” “source available,” and “community license” are emerging to describe the spectrum between truly open and proprietary.<\/p>\n The practical impact: when someone says a model is “open source,” always check the actual license. The label is unreliable.<\/p>\n For commercial products<\/strong>: Apache 2.0 models first. Zero legal ambiguity, maximum freedom, simplest compliance. Mistral and many Qwen models live here.<\/p>\n For research and experimentation<\/strong>: license matters less. Use whatever model performs best. Most research use cases fall within even the most restrictive model licenses.<\/p>\n For startups planning to scale<\/strong>: be cautious with user-threshold licenses like Llama’s. You probably won’t hit 700M users, but investors and acquirers will ask about licensing risk during due diligence.<\/p>\nThe Llama Community License<\/h2>\n
Qwen License and the China Factor<\/h2>\n
Other Licenses in the Wild<\/h2>\n
The “Open Source” Definition Debate<\/h2>\n
Choosing Based on Your Needs<\/h2>\n