Falcon 180B, open source AI and control over compute

October 25, 2023

In September, the Technology Innovation Institute from the United Arab Emirates released Falcon 180B, a new large language model that improves on the previous Falcon 40B model. The model, described by its creators as a “super-powerful language model,” is currently the top-ranked model on Hugging Face’s Leaderboard for pre-trained Open Large Language Models. Compared to the most popular closed models, it ranks second behind GPT 4, on par with Google’s PaLM 2 model. It is also one of a few examples of robust models that are publicly funded.

This opinion takes a closer look at how the Falcon 180B model is licensed and is a part of our exploration of the emergent standards for the sharing of AI models. Previous pieces in this series analyzed the RAIL licensing model, looked at the type of licensing used for models shared on Hugging Face, and studied the Llama 2 license.

The Llama 2 release, over the summer, came at a time of an increasingly heated political debate about closed and open AI development. This debate frames the issue of open source largely as one about AI risks, as demonstrated by conversations in the UK ahead of the AI Safety Summit. Another line of debate considers the competitiveness of open-source AI development. Some experts argue that open source development is “GPU poor” and cannot compete with the products of the largest company creating closed models.

Is Falcon an example of AI openwashing?

The Falcon release, just like that of Llama 2, adds to the confusion of what exactly is open-source AI. The Open Source Initiative is conducting a process to collectively define this, considering the specificity of machine learning (AI) technologies. In the meantime, AI labs release models described as “open source” or “open access” under conditions that limit sharing and use. The recent release of Llama 2 is a prime case of openwashing: it has been presented as “open source,” Meta describes it as “open source,” yet the license includes anti-competitive measures that break established norms around open source and don’t provide for the freedoms that open source secures.

Falcon’s release follows a similar pattern, where a narrative on open-sourcing the model goes hand in hand with licensing requirements that break the open-source standard. The TII institute states that all its models are “open source or open access” and describes the newest Falcon 180B model as “an open access model for research and commercial use.” The use of the term open access in this context is unusual and confusing, as Open Access refers to an open publication strategy for academic articles and journals. Even when used in a broader sense, it is usually used to refer to academic outputs.

The initial Falcon model, the Falcon 40B, was released under an Apache 2.0 software license compliant with the open-source definition. The new model, Falcon 180B, is released under a bespoke license based on Apache 2.0 but modified to include behavioral use restrictions and limitations on hosted use. The license is similar to Meta’s bespoke Llama license, which combines traditional open-source provisions with behavioral use restrictions and anti-competitive clauses.

While both types of modifications mean that a license is no longer compliant with the open-source definition, they must be considered separately. Most prominently introduced by the RAIL project, behavioral use limitations are a vital licensing innovation that aims to address some of the risks associated with AI development. While some proponents of open source see them as unacceptable limitations, others consider them to be ways of balancing sharing with risk management. And it is the anti-competitive clauses that are more controversial.

Falcon license and control over compute

These provisions aim to control the use of the Falcon model by cloud hosting providers by requiring them to obtain a dedicated license from TII. The Falcon 180B TII License Version 1.0 includes a section on “Hosting use,” which defines such use as:

 “any use of the Work or a Derivative Work to offer shared instances or managed services based on the Work, any Derivative Work (including fine-tuned versions of a Work or Derivative Work) to third party users in an inference or finetuning API form.” 

The provision does not cover uses of the model in applications or integrated end-user products. As such, it is focused on a narrow, but crucial category of commercial users: cloud providers. These companies run machine learning models as services, either as models that can run inference or as APIs that allow the models to be finetuned. As such, cloud-based models are an efficient way of deploying machine learning technologies by other entities. But as a result, power will concentrate in the hands of a handful of companies running these cloud services.

The fact that TII introduced this limitation shows how access to compute is a crucial issue that needs to be taken into account when considering how an open-source approach impacts AI development. The authors of the recently published “Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI” paper argue that the concentration of computational resources is one of the primary obstacles to open source AI development.

This argument can be extended, as access to computational resources is also necessary to run inference, that is, to deploy models as services available to users. We are now observing not just AI development happening at a frantic pace but just as rapid deployment, across all spheres of life, and types of organizations. All these services, including those that are public or even civic in nature, ultimately depend on an oligopoly of cloud providers. As a case in point, Wikimedia — seeking to offer new channels of access that benefit from machine learning — chose to start with a plug-in for ChatGPT.

As my colleagues recently pointed out,

“public institutions have been held back from updating their public interest missions in response to the challenges and opportunities resulting from digital transformation and (…) there is a need for more investment into public digital infrastructures. If we want to reduce our dependence on big tech, we — in the European context, that means the EU — must invest in public infrastructures. In the context of AI, this means establishing a robust and publicly accessible computational foundation for open source AI research.”

The introduction of anti-competitive provisions into the Falcon license confirms the argument from the “Open (For Business) Paper”: that access to compute is a crucial factor that will impact open source AI. The authors of the paper argue that control over computing resources, concentrated in the hands of just a few monopolistic companies creates a dependency for any open source project – as they have to rely on this hardware layer. As a result, the sharing of code does not democratize technology, but only brings new clients to the cloud companies.

One important issue that sheds more light on this licensing strategy is the fact that Falcon 180B was trained on Amazon’s cloud infrastructure and is now available through AWS cloud services. While there is no publicly available information on this partnership, AWS does not seem to be bound by the anti-competitive measures of the Falcon license. The Falcon licensing provisions can therefore be seen as protecting this relationship – and strengthening the position of Amazon.

The Falcon model and license case also show the importance of understanding the practical impact of opensourcing AI models on AI development and the specific ways in which open source models are used (e.g., through finetuning). For instance, the fact that the Falcon license prohibits the deployment of the Falcon model as a cloud service significantly limits its business uses – but most probably does not affect its use by researchers and amateurs, able to finetune and run inference on models they host themselves.

Just as with Llama 2, Falcon’s documentation does not explain the specific limitations added on top of the canonical open-source release model. And just as in the case of the previous model, the new release strategy is interesting in terms of understanding modifications to the open source approach happening in the AI development space. At the same time, it is potentially dangerous, as it dilutes the open source standard.


I am grateful to Aviya Skowron and Hailey Schoelkopf from Eleuther.ai for their advice on how Falcon licensing conditions affect open-source AI development.

Alek Tarkowski