Moonshot AI released Kimi K3 with no keynote and no model card, just a quiet overnight flip of kimi.com and its API. K3 is a sparse mixture-of-experts model with roughly 2.8 trillion total parameters, a 1-million-token context window, and native vision support, positioned for long-horizon coding and agentic workloads rather than short chat exchanges. On Artificial Analysis's composite leaderboard, K3 scored an Elo of 1,547, a jump of 732 points over its predecessor Kimi K2.6, placing it behind only Claude Fable 5 among all models tracked. On Moonshot's own benchmark comparisons, K3 mostly beats Claude Opus 4.8 max and GPT-5.5 high but trails Claude Fable 5 and GPT-5.6 Sol, putting it solidly in the second tier of frontier models rather than at the absolute top, while remaining dramatically cheaper: API pricing is $3 per million input tokens and $15 per million output tokens, well below what the top-tier proprietary models charge. Full open weights are promised by July 27, meaning K3 will join DeepSeek V4, Qwen3.5, and GLM-5 as another top-five open-weight model from a Chinese lab, reinforcing the pattern where open-weight models now trail the closed frontier by a matter of weeks rather than the year-plus gap that used to be typical. One illustrative data point that circulated widely: asking K3 to generate an SVG of a pelican riding a bicycle, an informal benchmark popularized by developer Simon Willison for testing spatial and compositional reasoning, took 16,658 output tokens, including 13,241 reasoning tokens, at a cost of about 25 cents, showing both the model's heavy reliance on chain-of-thought reasoning and the real per-query cost that comes with it even at K3's low headline pricing.