FROM MODEL TO APPLICATION

OPEN-WEIGHT MODELS VERSUS CLOSED, HOSTED MODELS The numbers inside a trained model are called its weights. An open-weight model is one whose weights the creator has published, so anyone who meets the license terms can download them and run the model on their own computers. A closed, hosted model is one whose weights are kept private, and you reach it only through the provider's API. Notice that open-weight is not always the same as open source. Open source usually means that the code and often the training information are shared under a license that allows broad use, while an open-weight release may share only the finished weights, sometimes with usage restrictions. Read the license before you build a business on one. Here is how the choice affects builders. With an open-weight model that you run yourself, you control your data, because nothing leaves your servers. You control your costs at large scale, because you pay for hardware and electricity rather than per request. You can customize the model by training it further on your own examples. But you must manage the hardware, keep the software updated, handle capacity and outages, and hire the skills to do so. With a closed, hosted model, you avoid hardware management and often get the most capable models available, with fast access to new versions. But you depend on the provider's uptime, prices, and policies, your data is sent outside your walls, and the provider may retire or change a model, which can alter your application's behavior. Neither choice is always right. A strong harness is designed so that the model can be swapped with limited effort, which is a major reason for the model adapter.