Late on the night of July 27, Moonshot AI, headquartered in Haidian, Beijing, released a statement unveiling the model weights and technical report for Kimi K3, and open-sourcing the key Infra (critical underlying infrastructure) technology that supports Kimi K3 model training. "Now, everyone can download and deploy the Kimi K3 model—whether for internal R&D or embedding into end-user products, it can be used freely."

Moonshot AI stated that Kimi K3 is its most capable model. The parameter scale of Kimi K3 is approximately three times that of Kimi K2.5, but scaling is not merely about stacking parameters. Under constrained computing conditions, the technical team leveraged a series of innovations—including Kimi Delta Attention, Attention Residuals, and MoonEP—to improve scaling efficiency by 2.5 times. In other words, in the sense of optimal computing resource utilization, the intelligence output per unit of computing power is roughly 2.5 times that of before.

Alongside the Kimi K3 model weights, Moonshot AI's model training methodology—the Kimi K3 technical report—has also been released simultaneously.

When asked why they chose to open-source their strongest model, Moonshot AI stated that they firmly believe in the value of open-weight models. These models lower the barrier to accessing intelligence, drive innovation, and grant users greater control over data, privacy protection, and ownership. Over the past few weeks, they have heard support from many in the AI community and technology leaders with a shared vision. "We believe that for a technology with such profound impact as AGI, a broad, open ecosystem is the most suitable soil to nurture it. To this end, we will continue to contribute our efforts."