OpenAI’s GPT-3 model had 175 billion parameters when it was introduced in 2020.
GPT-3 was described in the research paper “Language Models are Few-Shot Learners,” published in 2020. It used the Transformer architecture and was trained to predict the next token in text. The largest GPT-3 version contained 175 billion adjustable numerical parameters, making it dramatically larger than earlier widely discussed language models.
The model demonstrated few-shot learning: it could often perform a task after receiving a short set of examples in the prompt, without changing its parameters. Its abilities included generating prose, answering questions, translating text, and producing code, although it could also produce fluent false statements. Parameter count measures model capacity, not a complete score for intelligence, factual reliability, or usefulness. GPT-3’s release helped popularize large language models and in-context prompting.