Token streaming delivers a model's output incrementally as it generates: words appearing live rather than after a long wait. Perceived latency collapses: users
Delivering a model's output incrementally as it generates, words appearing live rather than after a long wait. Perceived latency collapses: users see progress immediately even when full responses take many seconds.
It is a real engineering surface: server-sent events or websockets, partial-output rendering, graceful interruption, and careful handling of structure, since streaming valid JSON requires care. It is the default UX of AI products.