Chip speculative decoding comparisons

69 speculative decoding comparisons across DeepSeek V4 Pro 0813 1.6T, DeepSeek R1, Kimi K3.1 Pro 2.8T, Kimi K3 2.8T, GLM 5/5.1, MiniMax M3 428B, and Qwen 3.5 397B-A17B. Each page compares inference with the speculative decoding method (MTP, EAGLE and others) enabled versus disabled on the same model and chip, covering throughput, cost and interactivity at matched operating points.

Kimi K3.1 Pro 2.8T

1 speculative decoding comparison with benchmark data on Kimi K3.1 Pro 2.8T