B300: FP4 vs INT4 Precision Comparison

Throughput at 54 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T (B300): FP4 hits 1059 tok/s/GPU, INT4 hits 403. Per-million costs land at $0.61 and $1.63 respectively. FP4 is 166% cheaper per token; FP4 delivers 163% more tok/s/GPU. The cost-throughput tradeoff from lower precision is only part of the picture — see the evaluation page for accuracy data.

Around the middle of the 35–111 tok/s/user interactivity band, at 73 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T (B300): FP4 runs 926 tok/s/GPU at $0.70/M tokens, INT4 runs 285 at $2.28/M. FP4 is 225% cheaper per token; FP4 delivers 226% more tok/s/GPU. Precision changes affect both inference speed and model quality — consult the evaluation tab for accuracy benchmarks.

At 92 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T (B300), FP4 delivers 911 tok/s/GPU at $0.71 per million tokens; INT4 delivers 193 tok/s/GPU at $3.34. FP4 is 367% cheaper per token; FP4 delivers 373% more tok/s/GPU. Lower-precision quantization trades model accuracy for throughput — check the evaluation page for quality impact. (Numbers reflect the default 1k/1k selection for this URL — table and chart below update if you change sequence or model in the controls. Each side uses the best available serving configuration for that precision, which may include speculative decoding such as MTP where recipes exist — the same convention as the other comparison pages.)