← All questions
Inference & ServingMedium· Latency· User Experience· Metrics

What’s the difference between TTFT and inter-token late...

TTFT (Time To First Token) scales with prompt length and model size, while ITL (Inter-Token Latency) scales with model weight bytes and batch KV cache size. In a streaming chat product, ITL dominates the user experience; if the text appears sluggish, the user perceives the system as slow.

While a 200ms TTFT is generally invisible to users, a 100ms ITL creates a perceptible lag. Therefore, optimization strategies differ: ITL is best improved via quantization and KV cache reduction, while TTFT is best addressed through multi-GPU prefill or prompt pruning.

Code playground (Python)