The ticket to the second half of the large model era is no longer just about stacking computing power, but rather a restructuring of underlying logic.

At the NVIDIA GTC2026 conference held on March 18, Moonshot AI founder Yang Zhilin delivered a highly anticipated public speech. This was his first systematic disclosure of the core technology roadmap behind the Kimi K2.5 model, offering new insights into the evolution of large models in the "post-Scaling" era.

Yang Zhilin pointed out in his speech that to break through the limits of intelligence, it is necessary to completely restructure key technologies such as optimizers, attention mechanisms, and residual connections. He summarized the evolutionary path of Kimi into three key dimensions that work in synergy:

Token Efficiency: Reject resource idleness, pursuing an even more extreme computing efficiency ratio.

Long Context: Continuously deepen Kimi's long-term memory advantage to process massive-scale information.

Agent Cluster: The form of intelligence is evolving from individual combat to dynamic-generated "digital clusters."

In Yang Zhilin's view, the current scaling has evolved into finding scale effects in efficiency, memory, and automated collaboration. If the technical gains from these three dimensions can be multiplied, the model will unleash an intelligence level far beyond its current state.

According to previous release information, the Kimi K2.5 launched in early January this year already demonstrated this "all-around" capability. As the most powerful open-source model from Moonshot AI so far, it adopts a native multimodal architecture, achieving state-of-the-art (SOTA) performance in code and visual understanding, and also supports flexible switching between "thinking" and "non-thinking" modes, accurately adapting to agent task scenarios.

With Moonshot AI's technological secrets being revealed, the focus of competition in the large model arena is shifting from "parameter count" to "intelligence density." When agent clusters become the ultimate form of future intelligence, whether Kimi can achieve a leap forward under Yang Zhilin's "three-dimensional multiplication" logic has become a focal point of industry attention.