MoE-first architectures and domestic hardware drive China's open-source AI ecosystem (Kimi K2, MiniMax M2, Qwen3, DeepSeek R1)

AI Impact Summary

China's open-source AI momentum is shifting toward Mixture-of-Experts (MoE) architectures (e.g., Kimi K2, MiniMax M2, Qwen3, DeepSeek R1) to balance capability with cost across diverse hardware. The shift is accompanied by a rapid expansion into multimodal and agent-based workloads, with end-to-end tooling and edge-to-cloud coordination increasingly emphasized and a push to run models directly on domestic chips (Huawei Ascend, Cambricon, Kunlun P800). Licensing is moving toward permissive Apache 2.0/MIT, and there is a focus on aligning model releases with inference frameworks, serving engines, and production-grade deployment (e.g., Mooncake serving, FastDeploy 2.0, Qwen ecosystem). Enterprises should plan for hardware-aware deployment, MoE routing strategies, and domestically provisioned compute to sustain scale and cost efficiency.