Model Marketplace

A model marketplace is a website where finished AI programs are collected and offered. Such programs have previously learned through an elaborate computing process from vast amounts of data, for example how to write texts or recognize images. This very learning process is expensive and often takes weeks. On a marketplace, one can instead simply take over the result, similar to how one downloads apps from an app store instead of programming them oneself. Some offerings are free and freely usable, others cost money per request. Each entry usually includes a description, sample code, and information about what the program is intended for.

Why hardly anyone starts from scratch anymore

Training a large language model from the ground up costs millions of euros in computing time, depending on size. Only a handful of corporations can afford this sum. For everyone else, model marketplaces are the only realistic access to modern AI. A small company downloads a suitable model and adapts it with a few thousand of its own examples. A project that used to take years thus becomes a project of weeks.

For the providers themselves, the platforms are a distribution channel. Anyone who publishes a new model immediately reaches millions of developers there. This explains why Meta, Mistral, or Alibaba make their models available openly: recognition and distribution are worth more to them than direct sales.

At the same time, a dependency arises. Anyone who bases their product on a rented model is dependent on its price and availability. If the provider changes the terms or discontinues a model, the product itself must be rebuilt. That is why companies take care to build their software so that the model can be swapped out.

What lies behind an entry

Technically, a model consists of a very large file of numbers, the so-called weights. These numbers are the result of training and determine the behavior of the program. A marketplace stores these files and makes them available for download. Added to this is a description page, often called a Model Card, with information on training data, limitations, and permitted uses.

There are two ways of use. With download, the model runs on one’s own computers, which brings control but requires expensive graphics cards. With the second approach, the model stays with the provider, and one sends requests to it via an interface. Billing is then based on the amount of text, measured in tokens, i.e., word units.

Also important is the license, i.e., the legal permission to use it. Some models may be freely used commercially, others only for research. The often-heard term open source does not always apply, since many providers release the weights but not the training data.

From Hugging Face to Bedrock

The best-known platform is Hugging Face, with over a million published models. It works similarly to a software library: anyone can upload something, and users rate and discuss it. In addition, the major cloud providers operate their own marketplaces, such as Amazon Bedrock, Azure AI Foundry, and Google Vertex AI. There, the focus is on vetted models that can be directly integrated into existing corporate IT.

These platforms regularly appear in business news. Reports such as “Model X is now available on Bedrock” mean that a provider is reaching new customers. Hugging Face’s valuation, most recently in the billions, is also read as a barometer for the AI market.

Anyone who wants to try it out themselves doesn’t need a company behind them. Many models can be tested directly in the browser, without installation. A common misconception, however, is that the largest model is automatically the best. For many tasks, a small specialized model is enough, one that responds faster and costs significantly less.