Moonshot AI released Kimi K3 on July 16th: a 2.8 trillion parameter Mixture-of-Experts model with open weights dropping July 27th. It ranks number 2 on the Vals AI Index, number 3 on Artificial Analysis's Intelligence Index behind only Claude Fable and GPT-5.6 Sol, and number 1 in Frontend Code Arena. It does this while being cheaper than the closed models it beats. The American-to-Chinese frontier gap, previously estimated at 6 to 9 months, is now closer to 3 to 5 months.
This is not a distillation story. K3 is Moonshot executing on core scaling fundamentals: data, algorithms, architecture, tools, and environments. The current frontier ranking puts Anthropic first, OpenAI second, and Kimi K3 third, ahead of xAI's Grok 4.5, Meta's Muse Spark 1.1, and Google DeepMind's Gemini Flash 3.5. Moonshot is doing this with substantially less compute than OpenAI or Anthropic, and the author notes that Kimi researchers were genuinely shocked to learn individual OpenAI researchers routinely access thousands of H100-equivalent machines.
The original piece is worth reading in full for two reasons. First, the author visited Moonshot's offices in China and offers a first-hand account of their engineering culture and resource constraints that explains how this result happened. Second, the analysis of what a confirmed open-weights K3 means for the broader ecosystem, versus a world where China has equivalent closed models, changes the strategic picture for every lab on that ranking list.
[READ ORIGINAL →]