• China’s open‑source AI community has pivoted to Mixture‑of‑Experts (MoE) architectures for cost‑effective scalability. • MoE allows dynamic compute allocation, enabling models to run efficiently across diverse hardware environments. • Domestic hardware adoption rises as companies prioritize local chips to meet regulatory and performance needs. • Open, reproducible models like DeepSeek R1 demonstrate that high reasoning can coexist with openness. • Licensing permissiveness and small‑model popularity reflect a shift toward broader community collaboration. • The “DeepSeek Moment” spurred competition, driving innovation in architecture, modality diversification, and deployment strategies.

Article Summaries:

  • China’s open‑source AI community is shifting its focus from raw performance to sustainable, flexible architectures. Since the “DeepSeek Moment” in early 2025, most leading Chinese models-Kimi K2, MiniMax M2, Qwen 3-have adopted Mixture‑of‑Experts (MoE) designs, enabling dynamic compute allocation and lower inference costs. Parallel to this, the ecosystem has expanded beyond text into multimodal and agent‑based domains, with StepFun, Tencent, and others releasing high‑performance audio, video, and 3D models and accompanying toolchains. A trend toward smaller 0.5‑30 B‑parameter models and domestic hardware adoption reflects a strategy prioritizing cost‑effective, widely deployable AI systems.

Sources: