“On the SVG Pelican benchmark, KIMI K3 consumed 13,241 reasoning tokens to produce a 3,417-token response.”