GPT-2

GPT-2 is a language model from OpenAI. Its knowledge cutoff is 2017-12-01.

GPT-2 is the 124-million-parameter OpenAI GPT-2 checkpoint tracked by the openai-community/gpt2 model card. It is part of OpenAI's second-generation autoregressive language model family, released in February 2019, and uses a decoder-only transformer architecture trained on WebText, a corpus assembled from outbound links posted on Reddit. This row is the compact 124M GPT-2 entry with a 1,024-token context window, not the larger 355M, 774M, or 1.5B GPT-2 sibling checkpoints.

GPT-2 is a pure language model: it predicts the next token given a context, without instruction tuning, RLHF alignment, or safety filtering. Users interact with it through prompt continuation rather than explicit instruction, which means it produces text stylistically consistent with its training corpus rather than executing user commands. Training data has a knowledge cutoff of approximately December 2017. The model does not support tool use, function calling, multi-modal input, or structured output.

GPT-2 is primarily of research and educational interest today. It established key patterns for large-scale pretraining and demonstrated emergent zero-shot task performance at scale, but the 124M checkpoint is substantially outperformed on practical tasks by later instruction-tuned models and by the larger GPT-2 variants. The model and its weights are available under an MIT-equivalent license on Hugging Face and via Azure ML. For new applications, GPT-2 is useful as a lightweight research baseline or for small text-generation experiments where the absence of alignment is acceptable.

GPT-2 has a 1k-token context window.