Kimi K3: The open-weights escalation

Moonshot AI released its latest flagship model, Kimi K3, on Thursday, July 16th. K3 is a 2.8 trillion parameter mixture-of-experts model, and its weights are scheduled for release on July 27th. Much of the accompanying analysis is written under the assumption that Moonshot keeps that weights-release date; the author notes that results land in a middle ground if the model is never released, and describes the view taken as a more extreme equilibrium.

The reported key fact is that either the open-to-closed or American-to-Chinese model performance gap has been reduced from a debated 6-9 months to something shorter, around 3-5 months. The release materials indicate K3 is a true frontier model, and the author states it will be the closest open models have been to the frontier since DeepSeek R1. Unlike R1, which the author characterizes as a Chinese lab pivoting quickly to reasoning models, K3 is described as executing on scaling known areas: data, algorithms, architecture, tools and environments.

Kimi K3 is reported at #2 overall on the Vals AI index, #3 overall on Artificial Analysis's Intelligence Index, behind only Claude Fable and GPT-5.6 Sol Max while being cheaper, and #1 overall in Frontend Code Arena. The author calls it the strongest open model ever released and says Moonshot AI is competing with Anthropic and OpenAI with far fewer resources. The piece argues that if adversarial distillation from U.S. closed frontier models contributed, it did so at most to a relatively small degree.

The author lists the state of peak model performance as: Anthropic's Claude Fable 5; OpenAI's GPT 5.6 Sol; Moonshot AI's Kimi K3 (open weights); SpaceXAI's Grok 4.5; Zhipu's GLM 5.2 (open weights); Meta's Muse Spark 1.1; DeepMind's Gemini Flash 3.5; and Alibaba's Qwen 3.7 Max, with Qwen 3.8 announced as also open-weights.

The piece also discusses Xi Jinping's keynote at the World AI Conference committing China's AI ecosystem to open-source and global diffusion, and covers technical details including Kimi Delta Attention, Attention Residuals, and a Stable LatentMoE framework activating 16 of 896 experts, cited as yielding roughly a 2.5x improvement in scaling efficiency versus Kimi K2. It notes Moonshot AI paused new subscriptions for K3 access while the API remains live, and mentions reporting attributed to Axios on past U.S. consideration of restricting Chinese open-source models.