Understanding Open Model Licenses
And what to look for in custom licenses
Google recently released Gemma 4, their next iteration of their small, open models with multimodal capabilities. The release was met with a lot of praise for the model’s performance, but more importantly, Google was praised for switching from a custom model license to releasing under the Apache 2.0 License. This greatly widened the potential applications for Gemma 4.
The positive reception to Google’s move highlights a crucial, often overlooked fact: model licenses are just as important as model capabilities. While many AI applications currently rely on closed APIs, the rise of powerful, accessible open models means developers must understand the licensing landscape to choose the right tool for the job and secure long-term rights.
Disclaimer: When I refer to ‘open models’ in the context of this article, I’m referring primarily to open-weight models and models that have released their weights and more.
Open models provide:
Cost Efficiency and Accessibility: Open models, especially smaller, efficient variants, can run locally or on specialized hardware.
Freedom from Vendor Lock-in: Permissive, open licenses grant developers permanent rights to use, adapt, and sell models. This protects developers from API changes, rate limiting, and changes to terms of use for models controlled by a separate entity.
Control and Customization: Full access to model weights enables developers to fine-tune models for highly specialized, complex reasoning tasks, MLOps, or agentic workflows.
Enhanced Security and Privacy: Open models can eliminate the need to transmit sensitive data to third-party API providers, ensuring data governance and privacy compliance.
Optimization and Predictable Latency: Open models enable developers to control the entire serving stack, allowing for custom performance optimizations and achieving predictable, low-latency inference critical for real-time applications.
Open models have already cemented themselves as a critical tool for building AI applications, but this is only becoming more true as the models mature. If you want to build AI applications, you need to understand the open model options available.
To understand those options, you must understand the licenses under which these models are released. Custom licenses can hamper the effectiveness of open models by explicitly prohibiting their use in certain applications and machine learning engineering tasks. Licenses, just as much as model capabilities, determine the right model for a job.
Below is some basic information on the most common open model licenses as well as what you need to look for when reviewing custom licenses. I’ve included some examples of models using each.
Common Open Licenses
The Apache 2.0 License
The Apache 2.0 License is a permissive open-source software license that enables developers to use, modify, distribute, and sell the software with minimal restrictions.
Additionally, it:
Doesn’t require developers to open-source derivative works of the software, meaning Apache 2.0-licensed software can be meaningfully integrated into proprietary enterprise software.
Explicitly grants patent rights to protect developers from patent litigation related to open-source code.
Requires developers to include the original copyright notice, license text, and a copy of any NOTICE files present in the open source code.
The Apache 2.0 License permits full commercial use of the software to which it is applied. Open models that are released under the Apache 2.0 License aren’t subject to any sort of prohibitive use policy or usage caps and can be freely used and adapted, including for commercial projects generating revenue.
Notable models released under this license are:
MIT License
The MIT License is one of the shortest and most permissive open-source licenses. It grants developers the right to use, copy, modify, merge, publish, distribute, sublicense, and sell software developed. It can be summarized by saying: “Do whatever you want with the code, just keep this license attached”.
While the MIT license is technically more permissive than the Apache 2.0 License, it doesn’t grant any explicit patent protection. The Apache 2.0 License is generally preferred for commercial use because of this. It ensures commercial software built on top of it isn’t at risk of patent infringement.
The most notable models released under this license are:
Custom licenses
Custom licenses can severely prohibit the usefulness of open models, but it entirely depends on what the custom license prohibits. The only way to understand custom licenses is by reading through them. When reading through them, you want to watch out for:
Scale Restrictions: Limits placed on commercial use, often defined by monthly active users (MAU), which can prohibit scaling applications.
Improvement Limitations/Derivatives: Strict prohibitions on using model outputs or derivatives (e.g., fine-tunes) to train or improve any other large language model.
Mandatory Attribution/Branding: Requirements to display specific branding or include the model’s name in derivative works.
Revocability and Unilateral Updates: Clauses allowing the releasing entity to unilaterally change the terms of use or revoke access to the model at any time.
Examples of custom licenses with these restrictions include:
Meta’s Llama Community License Agreement, which restricted the scale at which a Llama model could be used commercially (under 700 million MAU) and required developers to display “Built with Llama” in documentation and include “Llama” in the model name.
Gemma’s Terms of Use, which permitted commercial use but prohibited usage for harmful applications. The use policy could be changed by Google at any time, enabling Google to make a change and revoke usage rights at any time. This was changed with the most recent release, making Gemma models more universally useful.
Tongyi Qianwen License Agreement, which permits commercial use but requires authorization from Alibaba to exceed 100 million MAU. It also strictly prohibits the use of Qwen materials to improve any other large language model and requires developers to include “Built with Qwen” in product documentation.
While custom licenses are not inherently detrimental to a model’s utility—as seen with the early Llama releases that fostered entire ecosystems and naming conventions like Ollama—they do impact model adoption. The industry is currently witnessing a decisive shift toward more permissive, standardized licenses as developers increasingly prioritize ease of integration and legal certainty.
In summary, if you’re choosing an LLM for an application, you should:
Heavily consider open models for increased accessibility and flexibility.
Research both model performance and licensing.
Understand license limitations, especially for custom licenses. Prefer open licenses like MIT and Apache 2.0 for commercial applications.
Thanks for reading!
Always be (machine) learning,
Logan