As the large model industry enters a deep water zone of competing in practical applications and cost control, Moonshot AI's Kimi has increasingly clear commercialization paths. Recently, Huang Zhenxin, the B-end head of Moonshot AI's Kimi, shared the company's strategic thinking at an industry summit: to actively innovate in the underlying architecture rather than simply doing engineering stacking.

Regarding model pricing and business models, Huang Zhenxin emphasized that Kimi's positioning has always been a high-performance model. Although global computing power supply tension has increased model operation costs, Moonshot AI has effectively mitigated cost pressure through technical optimization, achieving a KV-Cache hit rate of over 90%, thereby providing users with cost-effective Token services. He clearly stated that evaluating model prices should not only look at the base pricing for input and output, but the cache hit efficiency in actual use is the key factor determining the user's final cost.

In terms of To B business layout, Kimi has shown a cautious attitude of "doing what it can and not doing what it cannot." Huang Zhenxin pointed out that Kimi will not get involved in heavy delivery businesses, but instead focus on continuous breakthroughs in model capabilities. The "last mile" customization services required for enterprise applications will mainly be completed by FDE (end-to-end) partners. Currently, Kimi has built a three-layer service system consisting of the underlying model, API architecture, and Agent products, and is deepening cooperation with industry giants like Amazon Web Services to promote solutions in vertical fields such as finance, healthcare, and manufacturing.

In terms of technology, Kimi shows obvious architectural orientation characteristics. The company has introduced the second-order optimizer Muon in training and launched the Kimi Linear Attention Architecture and Attention Residual Solution. These measures have significantly improved data usage efficiency, making the model more flexible when handling long text tasks. In response to the popularity of the "Harness" engineering, Kimi's internal team tends to practice "Loop Engineering," believing that as the base capabilities of models enhance, complex external engineering adaptation needs will gradually decrease.

Looking ahead, Moonshot AI will continue to focus on three dimensions: the intelligence level of agents, the ability to handle long context, and multi-agent collaboration. As high-performance models like Kimi K2.7 are gradually launched on cloud platforms, transforming energy efficiently into intelligence through technological innovation has become the core goal of this company in the long run of the AI industry.