Source: Tech Times — 2026-07-18
Summary
Two days after Kimi K3's debut, Moonshot AI's documentation now confirms the model exposes three reasoning-effort tiers — Standard, High, and Max — with explicit guidance on which tier fits which kind of task (casual conversation vs. debugging vs. hard reasoning/coding work). The tiering is live first in the Kimi Code environment; the main API quickstart docs haven't caught up yet.
Key Takeaways
- Reasoning-effort tiers let callers trade latency and cost against depth of reasoning per-request, rather than picking one fixed behavior for the whole model.
- Moonshot's docs explicitly map task types to tiers (conversational → Standard, debugging → High, hard coding/reasoning → Max), which is more prescriptive than most vendors' "just try it and see" guidance.
- The gap between Kimi Code's documentation and the main API quickstart suggests the tiering rolled out unevenly across surfaces, worth checking before assuming it's available wherever you're calling the model.
- This lands the same week Kimi K3 posted a near-frontier Terminal-Bench 2.1 score, so the effort-tier system is likely part of how it's competitive without burning max-effort compute on every call.