OpenAI’s largest released GPT-2 version had 1.5 billion parameters. OpenAI announced GPT-2 in February 2019 as a Transformer-based language model trained to predict the next word from internet text.
GPT-2 was released in several sizes, including models with 117 million, 345 million, and 762 million parameters, before the largest 1.5-billion-parameter version became available. OpenAI initially staged the release of the largest model because of concerns that large language models could be misused to generate convincing text at scale.
The figure is often confused with GPT-3’s 175 billion parameters, announced in 2020. Parameters are learned numerical values inside a model; they are not a direct count of words, documents, or facts. A larger parameter count can increase capacity, but it does not by itself guarantee accuracy or safe behavior.