Tokens Per Second (TPS)

Tokens Per Second (TPS) is the primary throughput metric for evaluating AI language model inference performance. It measures how many tokens a model generates per second after the generation process has begun. TPS and Time-to-First-Token (TTFT) jointly determine the overall user experience quality. A token roughly corresponds to 0.75 words in English or 0.5–0.6 words in other languages. Typical TPS benchmarks: Groq's LPU achieves 500–800 TPS for 7B parameter models; Anthropic's Claude API delivers 30–100 TPS depending on model tier; self-hosted open-source models on a single H100 GPU achieve 50–200 TPS depending on model size. TPS influences UX in two distinct ways. For short responses (up to ~500 tokens), TTFT dominates perceived responsiveness. For long outputs — documents, code, analyses — TPS becomes the determining factor. At 30 TPS, generating a 3,000-word document takes ~80 seconds; at 200 TPS, ~12 seconds. For voice AI systems, a minimum TPS of 100 is necessary for speech synthesis without perceptible gaps. Factors affecting TPS: model size (larger = lower TPS per request), quantization level (FP4 > FP8 > BF16 in throughput), batch size (larger batches increase aggregate TPS but lower individual TPS), hardware, and KV-cache utilization patterns.