MoE, Tiny Models, and Domestic Chips: China's Open-Source AI Playbook After DeepSeek

The open-source flood that followed DeepSeek’s breakthrough in January 2025 wasn’t just about bigger models. It forced a rapid, pragmatic convergence in how Chinese AI labs build and deploy technology, and the architectural choices are now unmistakable. Mixture-of-Experts (MoE) isn’t a research curiosity anymore—it’s the default. Models like Kimi K2, Qwen3, and MiniMax M2 all use it. The logic is brutally practical: MoE acts like a controllable compute distribution system, activating different numbers of ‘experts’ per task so a single deployment doesn’t need to burn a full set of resources for every query. It’s the engineering answer to real-world constraints, prioritizing sustainable operation and flexible deployment over chasing the absolute highest benchmark score.

Beyond the text-based race, the scramble for multimodal supremacy kicked into high gear. The community stopped just shipping model weights and started releasing entire engineering stacks—inference toolchains, distillation datasets, and agent frameworks—for video, 3D, and audio. StepFun’s latest speech-to-speech model, Step-Audio-R1.1, claims state-of-the-art performance that beats proprietary alternatives, while Tencent’s open-source Hunyuan models are pushing competition into video and 3D generation. This isn’t a collection of isolated stunts; it points to reusable, system-level capabilities being built in the open.

A quiet but significant shift is the preference for smaller models. The 0.5B to 30B parameter range is where the real action is for developers who need models they can run locally, fine-tune cheaply, and integrate into business systems. Qwen 1.5-0.5B has the most derivative models, highlighting that practical utility often trumps raw scale. The large MoE behemoths are increasingly acting as ‘teacher models’ to distill capabilities into these smaller, deployable workhorses. This is further lubricated by a decisive move toward permissive Apache 2.0 licenses, slashing the legal friction for companies to actually use these models in production.

The final piece snapping into place is the hardware. Model releases in 2025 are aligning tightly with inference on domestic Chinese chips, moving the ecosystem from a model-first to a hardware-first posture. DeepSeek didn’t just open-source a reasoning model; it proved that advanced reasoning could be engineered and reproduced openly, which directly inspired overlapping competition and a heavier focus on building a self-reliant stack from silicon to software.