Kimi K3 launched July 16 as a 2.8-trillion-parameter Mixture-of-Experts model — roughly 75% larger than DeepSeek's V4 Pro — with a 1-million-token context window and native multimodal input, landing at the frontier tier on several independent benchmarks while trailing Claude Fable 5 and GPT-5.6 Sol on broad measures. At launch, the model exposed only a single reasoning_effort setting: max, always on, with reasoning tokens billed as output at $15 per million tokens.
That single-tier design proved expensive fast: independent testing found a simple SVG-generation prompt consuming over 13,000 reasoning tokens to produce roughly 3,400 tokens of actual output, costing around $0.25 for a trivial request. Moonshot's updated Kimi Code documentation, published July 18, now maps all three effort levels — Standard, High, and Max — with the rollout reaching the Kimi Code environment first; the main API quickstart still lists only max as supported, suggesting a staged release.
The addition gives developers their first real lever on the speed-versus-compute tradeoff for a model Moonshot has positioned as evidence that open-weight models are closing the gap with the closed Western frontier — full open weights for K3 are dated for July 27.