AI EDUCATION: What Is Open-Weight AI?

We tend to think of openness as being good or having utility. You can patronize an open business. You can walk through an open door. You can trust someone who is an “open book.” 

But it’s not so open-and-shut when it comes to AI, especially in the financial services sector. 

Welcome to AI Education, where we’re going to discuss the topic of openness in AI, specifically, what it means for AI models to be open weight and/or open source. We don’t think we can get into this discussion without first talking about openness in software development in general, which is why, in a column where we meant to focus on open-weight AI models, we’re instead first going to tackle the open-source question and openness as a more general concept. 

Free software would be the least specific descriptor we need today—free software is released without charge to the user or customer, and can be used however they want. Open-source software, on the other hand, is released under a license that allows users to use, change and distribute the software and its source code to anyone for any purpose. It’s not just free, the most important aspect of open-source is that its source code can be used by anyone for any purpose, be it commercial or non-commercial. Open source came from the early years of computing where developers shared programs and pieces of code with each other and openly published the underlying code to their products to encourage innovation, a practice that was only somewhat curtailed as software became commercialized and shifted towards more proprietary models. 

What Is an Open-Weight AI Model? 

We need to first understand what is meant by weight in the context of an AI model.  A generative AI model answers questions by reading a question, retrieving and restating pertinent data and synthesizing a response. Generative AI learns how to do this by forming numerical parameters, called weights, that determine how a user input influences the software’s output. Weights are not programmed into an AI model via code, but are learned during its training process and, in some cases, are intended to continue to change over time as a model is used.  

In an open-weight AI model, the model’s developers have decided to put their cards on the table so that users and other developers can see what—and usually how—a model has “learned” during its training. Furthermore, open-weight models can, at least theoretically, downloaded and run on an institution’s own computers or even edge devices. 

Usually, when we say a model is open weight, it means that users can not only see the trained parameters, but are also able to adjust them to modify, specialize and otherwise build upon the model, just as open-source software usually implies that not only is the source code of software available for public view, but it can also be used and modified as needed. Yet many open-weight models omit the training data used to arrive at their weights—so in some ways, they are less transparent than proprietary models. 

The benefits of open-weight models are not only found in innovation. Open-weight models permit users the ability to understand a model’s mechanics, eliminating a potential black box around the software—which is something the financial services industry greatly desires. On the other hand, open weights also mean fewer controls around how AI is used and what it ultimately produces, which is something that highly regulated industries probably want to avoid—and it’s unlikely that any financial company would use an AI model without access to the material used to train it. 

What Is Open-Source AI? 

An open-source AI is an open-weight AI with the training data included—so users and institutions can not only access and modify AI systems and models, but any related tools, the source code, model weights and training data. Furthermore, any user is free to distribute the results of any modifications they’ve made to the AI system or model without the need to ask permission of any other developer, property owner or group. 

Some of the AI out there purporting to be open source isn’t really open source. 

While there are many examples of open-weight AI models, relatively few AI models are fully open source. These include DeepSeek-R1, Falcon-7B, Granite, Pythia, Amber and Crystal. While DeepSeek-R1 was developed by the Chinese startup DeepSeek, and Granite came out of IBM, most open-source AI is developed by consortiums and non-profit organizations. 

Open-source models are being used for the same tasks that their proprietary brethren are performing, including large language model chatbots, translation and code generation, computer vision, robotics, healthcare diagnostics, drug discovery and defense applications. 

Why Isn’t There More Open-Source AI? 

For one thing, proprietary systems are often immensely profitable, but they suffer from stagnation—a lack of the innovation that usually occurs around open systems. But that hasn’t stopped the open-source movement elsewhere. Despite the many moves to fence off the internet and create closed-off systems, a sizable portion of the IT and developer community remains committed to openness and depends on openness for continued innovation. To some extent, technology companies have recognized this need and supported the open-source movement. 

Open-source models do come with some serious concerns around transparency. Ironically, while the weights are open and adjustable, and the training data is published for anyone to see, it’s not always easy to understand and explain the decisions made by open source models—in a way, they operate in a black box of their own making, limiting the ability of highly regulated industries to adopt open-source models. We need to be careful of any presumption or bias that tells us that more openness is always better. But something more holds open-source models back. 

The real reason there isn’t more open-source AI out there is the high price of intellectual property. AI models are trained on a massive blend of data which usually, somewhere, includes proprietary data or copyrighted information. The machines that publish this information cannot distinguish it from the non-proprietary data within a training set—so publishing the training data amounts to piracy.