The Black Box That Thinks in Code: Kimi K3's Memorized Index and the Cost of Closed-Loop Reasoning
Design Arena's teardown of Kimi K3's thinking traces surfaced something genuinely strange. Moonshot's open-weight model doesn't just reason more than its competitors — it uses over 12x the reasoning tokens of Claude Opus 4.8 and more than double its own predecessor, K2.6. The reason: K3 runs what is effectively a full agent loop inside its chain of thought. It plans, writes sample code for individual components, mentally tests interactions, iterates, and only then emits a final answer. More reasoning tokens are…