Deconstructing the Kimi K3 Technical Report: Moonshot AI Starts Setting Problems for DeepSeek as Two Scaling Axes Redefine the Frontier

Moonshot AI released a 47-page Kimi K3 technical report redefining AI scaling as both pre-deployment training compute and post-deployment inference compute.

Kimi K3 is said to reach 2.78 trillion total parameters with 104.2 billion active per token, more than doubling DeepSeek V4 Pro.

The report highlights three core innovations: KDA linear attention for sequence scaling, AttnRes for depth routing, and Stable LatentMoE for width scaling with formal load balancing.

It also introduces MoonEP for expert communication and a VM-based RL sandbox, while claiming strong results on browsing and coding benchmarks.

Overall, Moonshot AI is positioning Kimi K3 as an architecture-led challenge to DeepSeek and the next phase of frontier AI competition.