Kimi K3's design secret may lie in its Thinking Traces

Moonshot AI's latest open-weight model, Kimi K3, has taken 1st place on Frontend Arena with an Elo of 1392. This is a significantly higher score than its previous models, Kimi K2.6 and K2.7 Code.

The core secret behind Kimi K3's performance lies in its overwhelming use of Thinking Tokens. Kimi K3 goes through more than 12x the reasoning process of Claude Opus 4.8, and more than 2x that of Kimi K2.6.

Analysis shows that Kimi K3 goes beyond simple reasoning to simulate the following agentic workflow within its Chain of Thought:

• Multi-stage thinking steps: A step-by-step approach that progresses from planning to decision-making to individual component design
• Code-based design: Meticulously designing interactions by directly writing sample code during the thinking process, before producing the final output
• Code-centric reasoning: Unlike previous models, a substantial portion of reasoning tokens is allocated to actually writing code

Thanks to this unique CoT approach, Kimi K3 shows greatly improved capabilities in creative component design and integrating external dependencies.