Big Tech's Open-Weight AI Strategy and Its Business Logic

Artificial Intelligence models learn patterns from data during a one-time process called training. Running the trained model later to answer a question is called inference. The numbers a model learns during training are called weights. They shape how it processes an input and produces an output.

Most people reach AI models like GPT or Gemini through a paid app or an API (meaning: a channel letting one software system use another's service). These are closed or proprietary models. The company alone controls access. An open-weight model is different. It makes its trained parameters available for download. Any developer can run it on their own computers. They can change it or build it into their own product. This is not new. OpenAI released the weights of its GPT-2 model back in 2019. What has changed recently is how capable these downloadable models have become. Chinese developers have also started serious competition.

It is important to note that being open-weight does not automatically mean fully open-source. A company can share only the weights. It can keep its training data and training method secret. It can attach a licence that limits how the model may be used. For instance, it can bar commercial use.