Moonshot AI said Kimi K3, unveiled July 16, will publish its weights on July 27, potentially making it the strongest open model yet released.
Benchmark claims place K3 at No. 2 on Vals AI, No. 3 on Artificial Analysis and No. 1 in Frontend Code Arena, narrowing the open-to-closed or U.S.-China gap to roughly 3-5 months.
The 2.8T-parameter MoE model uses 16 of 896 experts and architectural changes including Kimi Delta Attention, which Moonshot says lifted scaling efficiency about 2.5 times versus Kimi K2.
The release lands as China doubles down on open-source AI—Xi Jinping backed open diffusion at WAIC—and as Alibaba signaled a 2.4 trillion-parameter Qwen 3.8 open-weights model is coming.
That combination could pressure the economics of closed U.S. labs by lowering margins for frontier models while accelerating broader AI diffusion and intensifying policy debates over open-weight risks.
As China offers powerful AI to the world for free, is Silicon Valley's expensive, proprietary model now obsolete?
Is China's open-source strategy a greater threat to US tech dominance than the race for advanced computer chips?
With Chinese AI allegedly built on stolen US tech, how can Washington protect its most critical intellectual property?
Kimi K3 and Qwen3.8: China’s Trillion-Parameter Open-Source AI Models Reshape the Global AI Landscape
Overview
The report highlights the groundbreaking release of Moonshot AI's Kimi K3, an open-weight large language model with 2.8 trillion parameters, making it the most powerful model ever released. Its open-source nature allows developers to freely download, modify, and deploy the model, which is set to disrupt existing commercial AI models, especially in Silicon Valley. Running Kimi K3 locally requires significant computing resources, but its accessibility is expected to accelerate global AI innovation. The model's capabilities have even raised concerns among US tech leaders about falling behind China, underscoring the growing impact of open-source AI on the global stage.