Large language models (LLMs) have become a focal point of excitement and speculation. With each new model release, particularly those labeled as “open source,” the AI community buzzes with discussions about their potential impact on the industry. Many draw parallels to the rise of Linux in the server industry, predicting a similar trajectory for open source LLMs. However, as a researcher working on large language models, I believe it’s crucial to examine this enthusiasm critically and dispel some misconceptions about what truly constitutes “open source” in the context of LLMs.

Open-source means you’re willing to develop the project out in the open.

The term “open source” has been liberally applied to many recently released language models, but this usage often misses the essence of what open source truly means. Open source is not merely about the end product being freely available or distributed under a permissive license. It’s a culture, a methodology, and a collaborative approach to software development. True open source projects are characterized by transparency in the development process, community participation and contribution, the ability to fork and modify the project, and a clear provenance of ideas and code. When we examine the so-called “open source” LLMs through this lens, we find that they fall short in several critical aspects.

Most LLMs labeled as open source are, in fact, the product of well-funded companies with significant resources at their disposal. These companies develop their models behind closed doors, using proprietary datasets and substantial computational power. The development process is opaque, with no real-time community involvement or contribution. What these companies release is not a truly open source project, but rather the final model weights, a research paper describing the training process, and sometimes, portions of the training code. While this information is valuable, it does not constitute an open source project in the traditional sense. The community has no insight into the decision-making process, the failed experiments, or the iterations that led to the final product. There’s no ability to contribute to the core development or to influence the direction of the project.

Indeed, one of the most critical aspects missing from the current “open source” LLM landscape is provenance. In a truly open source project, you can trace the evolution of ideas, see the debates that shaped decisions, and understand why certain approaches were chosen over others. This history is invaluable for replication, education, and further innovation. With current LLMs, we’re given a snapshot of the final product without the rich context of its development. This limits our ability to truly learn from and build upon these models in a foundational way.

It’s important to acknowledge the unique challenges that LLMs pose to the open source model. The sheer computational resources required to train these models make it impractical for most individuals or small organizations to meaningfully contribute to or fork the project. This technical barrier fundamentally alters the dynamics of what open source can mean in the context of LLMs.

The most open-source aspect of these LLMs actually occurs after their release. The community’s role is primarily in fine-tuning the models for specific tasks, developing tools and interfaces, optimizing the models for different hardware, and exploring novel applications. While this ecosystem of post-release development is vibrant and valuable, it’s important to recognize that it’s built upon a foundation that the community had no part in creating.

A crucial point often overlooked in these discussions is that there has never been a truly community-built LLM (except maybe for the Bloom project which failed). All significant LLMs to date have been created by well-resourced companies or institutions. This stark reality stands in sharp contrast to the collaborative, community-driven development that characterizes genuine open source projects like Linux. Rather than comparing these so-called “open-source” LLMs to Linux, a more accurate analogy might be to proprietary software that allows for extensive customization. It’s akin to building upon Windows or macOS, where you can create applications and modify certain aspects, but you have no control over or insight into the core operating system.