CHINA’S MOONSHOT AI DROPS THE WORLD’S LARGEST OPEN-SOURCE MODEL AND IT IS CLOSING IN ON GPT-5.6 Moonshot AI released Kimi K3 on July 16, and the numbers are staggering. The model carries 2.8 trillion parameters, making it the largest open-source AI model ever released. For context, most frontier models that reach this capability range are kept entirely proprietary. Moonshot is giving it away under a modified MIT license, with full weights available by July 27. Under the hood, Kimi K3 runs a Mixture-of-Experts architecture built on Kimi Delta Attention, a hybrid linear attention mechanism that activates just 16 of its 896 experts on any given task. It handles a 1 million token context window natively and processes text, images, audio, and video in a single pass. Priced at $3 per million input tokens and $15 per million output tokens, it undercuts most US frontier models while matching them on benchmarks. Scores from the Artificial Analysis Intelligence Index put Kimi K3 at 57, level with Anthropic’s Opus 4.8 and OpenAI’s GPT-5.5, behind Fable 5 and GPT-5.6 Sol but not by much. That gap is closing faster than anyone in the US expected two years ago. Chinese open-source models are no longer academic curiosities. They are production-grade competitors, and now anyone in the world can download one that rivals the best closed models America has built. Keywords: Kimi K3, Moonshot AI, open-source AI model, China AI 2026