Original model · Public beta

GPT-2 chat

A conversation inside a 1,024-token window.

GPT-2 XL · 2019 checkpoint

Leave Seed blank before New chat to choose a seed; it is shown and reused for each reply. Temperature 0 uses greedy decoding and ignores the seed. Example rounds are applied when you choose New chat.

Current model context

— / 1,024 tokens

The complete retained context: fixed header, seeded examples, and recent messages. Human messages are blue; GPT-2 replies are red. Older rounds leave this window as it fills.

Starting a chat…

Enter to send · Shift+Enter for a new line

Regenerate uses the same input prompt with the current temperature and reply limit. It replaces the latest reply and increases the seed by 1 (blank uses the chat seed); the maximum seed wraps to 0. Choose a temperature above 0 to sample an alternative.

Connecting to GPT-2…

Import and export the conversation and generation settings as JSON.

Your conversation and draft are saved in this browser when storage is available. Each tab keeps its own chat; reopening this page restores the most recently used chat. Export JSON to move it to another browser or keep a backup. The full transcript and its generation records are sent to the server when restoring, importing, previewing, or requesting a reply. One reply runs at a time; other visitors may be ahead in the queue.

GPT-2 completes text; it was not trained as a chat assistant. Its replies can be inaccurate, repetitive, or offensive.