Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
Total generation time is unchanged; perceived latency collapses. A response that takes twenty seconds feels broken delivered at once and acceptable delivered progressively, which is why nearly every chat interface streams.
It complicates anything that needs the complete response before acting — validation, structured parsing, moderation, tool dispatch. The usual resolution is to stream the prose to the reader while buffering the parts the system must see whole.