Three Major Technology Trends in China's Open-Source AI

Since the 'DeepSeek moment,' China's open-source AI community has been focusing on architectural strategies for sustainable and flexible deployment, going beyond just model performance.

Standardization of MoE (Mixture-of-Experts) Major models such as Kimi K2, MiniMax M2, and Qwen3 are adopting the MoE architecture as the default. This is a strategic choice to maximize cost efficiency within limited computing resources and to enable flexible deployment across diverse hardware environments.

Multimodal and Agent-Centric Expansion Moving beyond a text-centric focus, competition is expanding into various modalities including Any-to-Any, video generation, audio, 3D, and agents. StepFun's audio/video models and Tencent's Hunyuan Video are representative examples, with efforts focused on securing system-level capabilities encompassing inference, datasets, and workflows, beyond simply releasing models.

The Rise of Small Models and Knowledge Distillation Small models in the 0.5B–30B range, which are easy to run locally and integrate into businesses, are gaining significant popularity. Distillation, which uses large-scale MoE models as a 'Teacher Model' to transfer their capabilities to smaller models, has established itself as a core structure of the ecosystem.