On openness · The But Is It Open? blog

Apache 2.0 Is a License, Not a Community

· 4 minute read

Gemma 4’s Apache license is worth celebrating. But permission to use a model is not a path to help govern it—and open source once taught us to expect both.

Detail of the approved cover painting: layered burgundy, ochre and navy planes, diagonal strokes and scraped pigment on ivory.

Apache 2.0 Is a License, Not a Community

Google released Gemma 4 under Apache 2.0, and I’m glad they did. It’s a good open-source license, with familiar terms that let people use, modify, and redistribute the model, including commercially. Google didn’t attach a separate list of prohibited uses or a rule that changes the deal once your business gets big enough. There are ordinary obligations around license copies and notices, but no custom bargain to decipher.

I want more model developers to make that choice. It gives people real independence: you can download the weights and run the model without depending on Google to keep an API available.

But I still find myself wanting something more. When I see an Apache license, I tend to imagine a community behind it that I could join. With AI models, that assumption needs a closer look.

Something you could join

Part of what drew people to open source was the possibility of participating. You could use a project, find a problem, contribute a fix, and eventually earn some responsibility for its future. The Apache Software Foundation describes that progression as meritocracy. It was never perfectly fair, but the idea mattered: your contribution could give you a voice, whether or not the company behind the project employed you.

Apache 2.0 doesn’t require that arrangement. You can adopt the license without adopting the foundation’s way of working. Still, the two have enough history together that it’s easy to read the license and assume the invitation comes with it.

Gemma does offer ways to participate. Its public library welcomes pull requests, and developers build applications and fine-tuned variants around the model. That work counts. But improving the tools for using Gemma is different from helping decide what goes into the next model.

What went into it?

The weights let me run the result. They don’t tell me everything I’d like to know about how Google produced it. What was the training data? How was it selected and mixed? How could someone outside the team challenge a choice or contribute an improvement?

Google’s model card describes broad categories of training data and processing. That’s useful information, but it isn’t the same as being able to inspect the complete process or reproduce it. This is why I separate access, permission, and independent deployment from transparency and reproducibility. Gemma gives me much clearer answers to the first three.

I’m not suggesting every piece of training data can simply be published. Privacy and redistribution rights can make that impossible. But those limits leave questions that a license on the finished weights cannot answer.

Training may not work that way

I also don’t know how much of the older participation model can survive the economics of AI training. A maintainer can review a software patch and run its tests. Testing a change to a large model’s training data may require an expensive training run, with a coordinated team and substantial computing resources. Someone has to pay for that experiment before anyone knows whether it helped.

That makes “send us a patch” a much harder invitation to offer. Even with good intentions, Google may not be able to open the core training process in a way that resembles a conventional open-source project.

I can respect that difficulty and still notice the gap. We have a license associated with collaborative development on something whose production may not support the same kind of collaboration. There’s a community around the model, but less opportunity to become part of the team making it.

I’d like to know where outside work can influence the next release, even if the answer is limited. In the meantime, Google deserves credit for choosing Apache 2.0 without adding weird rules. It’s a great license. It just leaves me asking what more openness could look like here.

But is it open? For using and building with Gemma, the answer is encouraging. For understanding how it was made and joining in, I’m still left wanting more.

Derived from the openness research behind But Is It Open?

← More from the blogJoin the newsletter ↗