Closed-Weight Model

At its core, an AI model consists of a huge table of numbers. These numbers are called weights, and they contain everything the model absorbed during learning. Whoever owns the weights owns the model and can run it on their own computers. With a closed-weight model, the manufacturer does not release these numbers. You are allowed to use the model, but only through the company’s servers, for instance via a website or an app. The English word “closed” means geschlossen, “weights” means Gewichte.

Whoever has the weights has control

Training a large model often costs sums in the hundreds of millions. Companies like OpenAI, Anthropic, or Google protect this investment by withholding the weights. Their business model is ongoing access: customers pay per request or per month. If the weights circulated freely, anyone could run the model themselves and would no longer have to pay.

For users this has noticeable consequences. A company that uses a closed-weight model sends its data to foreign servers. For hospitals, law firms, or government agencies this is often a problem. Moreover, only the provider decides when a model is shut down, changed, or made more expensive. Anyone who has built their software on top of it then faces a rebuild.

Security researchers, however, also see an advantage. Secret weights mean that no one can simply strip out the built-in safety rules. With freely available models, such restrictions can be removed with manageable effort. Both sides have arguments, then, and the debate remains unresolved to this day.

Access only through the interface

Access almost always runs through an API. This is a defined interface through which a program sends a question to the server and receives an answer back. The user sees only these two ends. What happens in between remains hidden. You can picture it like a vending machine: you drop in a question, an answer comes out at the bottom, and the inner workings are sealed shut.

The size of such a model is also usually unknown. Companies rarely publish how many weights their system has or what texts it was trained on. Outsiders can only gauge the quality through test questions, not by recalculating anything. Experts call this a black box, meaning a sealed container with no insight into it.

An important distinction is one that often gets confused. The counterpart is called an open-weight model: there, you download the weights and run the model yourself. However, this is not the same as open source in the classic sense. The training data often remains secret, and the license restricts commercial use. Freely available weights therefore do not automatically mean full openness.

These models power almost all well-known chatbots

The best-known AI services work with closed weights. These include ChatGPT from OpenAI, Claude from Anthropic, and Gemini from Google. Many features in office programs, search engines, or phone operating systems also draw on such models in the background. So you often use them without even noticing.

In business news the term usually comes up in connection with competition. When a provider like Meta or Mistral releases a strong model with open weights, the pricing model of the closed providers comes under pressure. Analysts then say that the lead of the closed systems is shrinking. Such news regularly moves stock prices.

For you in everyday life, the difference shows up quite practically. A closed-weight model only works with an internet connection and an account with the provider. An open model can also run offline on a decent laptop. Once you’ve tried both, you understand the difference faster than through any definition.