On the evening of July 28, Moonshot AI officially open-sourced the 2.8 trillion parameter model Kimi K3 and released the model weights and technical report; on the same day, MoLeThread's official Weibo announced that it had completed the ultra-fast adaptation and stable startup of Kimi K3 based on its AI training and inference integrated smart computing card MTT S5000 and MUSA software stack.

As the first globally open-sourced large model at the 3 trillion parameter level and the most powerful version of Kimi so far, Kimi K3 features 2.8 trillion parameters and a 100,000 token context window, natively supports visual understanding, and integrates several architectural innovations such as the KDA hybrid linear attention mechanism (Kimi Delta Attention), attention residuals, Gated MLA, and Stable LatentMoE, specifically designed for cutting-edge intelligent scenarios such as long-range programming, knowledge work, and complex reasoning.

This ultra-fast adaptation not only demonstrates MoLeThread's MUSA software ecosystem's rapid response and full-stack collaboration capabilities for cutting-edge large-scale models, but also further verifies the solid capabilities of domestic full-function GPUs in supporting the research and development and engineering implementation of trillion-parameter large models. With the deep integration of high-performance domestic computing power and top-tier open-source ecosystems, the large model industry is accelerating towards an efficient, self-controlled, and fully integrated hardware-software innovation stage.