Streaming

What is streaming?

You've probably noticed it: when you chat with an AI, the answer appears line by line instead of arriving all at once after a long pause. That's streaming. The model generates and pushes results to you in real time — you see the answer while it's still being written, not a finished block.

How does it work?

It sends the answer one piece at a time
Generation is token-by-token. In normal mode the model collects a whole paragraph before returning it; streaming sends each token out the moment it's made, and the frontend renders them one after another.
It rides on chunked transfer
Under the hood, it usually uses SSE (server-sent events) or HTTP streaming to cut data into small pieces and keep pushing them, so the browser or app renders progressively.

What problem does it solve?

It kills the "waiting" anxiety
A long answer can take tens of seconds; a one-shot return feels like staring at a spinner. Streaming shows words moving immediately, slashing perceived latency.
It feels more like a conversation
Text appearing character by character feels like a human typing — the AI seems to be "thinking", not frozen.
You can bail early
If it's going off the rails, you can interrupt anytime instead of waiting for a whole block, saving time and tokens.

Anything to watch out for?

Streaming doesn't compute any faster — it just shows you sooner. It doesn't cut total compute time, but it crushes the feeling of waiting. For chat, code completion and long-form writing, where you want to watch it unfold, it's the default.

Bottom line: streaming shows you the answer as it's made, so an AI's reply unfolds like a human typing.