Architectural Choices in China's Open-Source AI Ecosystem: Building Beyond DeepSeek

This is the second blog in a three-part series on China's open source community's historical advancements since January 2025's "DeepSeek Moment." In this second piece we turn our focus from models to the architectural and hardware choices Chinese companies have made as openness becomes the norm.

For AI researchers and developers contributing to and relying on the open source ecosystem and for policymakers understanding the rapidly changing environment, architectural preferences, modality diversification, license permissiveness, small model popularity, and growing adoption of Chinese hardware point to leadership strategies across a multitude of paths. DeepSeek R1's own characteristics inspired overlap and competition, and contributed to heavier focus on domestic hardware in China.

In the past year, leading models from the Chinese community had almost unanimously moved toward Mixture-of-Experts (MoE) architectures, including Kimi K2, MiniMax M2, and Qwen3. R1 itself was an MoE model, it also proved a crucial point: strong reasoning could be open, reproducible, and engineered in practice. Under China's real-world constraints, maintaining high capability while controlling cost, and ensuring models could be trained, deployed, and widely adopted, MoE emerged as a natural solution.

MoE is like a controllable compute distribution system; under a single capability framework, compute resources are allocated across requests and deployment environments by dynamically activating different numbers of experts according to task complexity and value. More importantly, it does not require every inference to consume the full set of resources, nor does it assume that all deployment environments share identical hardware conditions.

The overall direction of Chinese open-source models in 2025 was clear: not necessarily the strongest possible performance, but ability to operate sustainably, deploy flexibly, and evolve continuously, achieving the best cost performance balance.

Starting in February 2025, open-source activity was no longer focused only on text models. It quickly expanded into multimodal and agent-based directions: Any-to-Any models, text-to-image, image-to-video, text-to-video, TTS, 3D, and agents all progressed in parallel. What the community pushed forward was not just model weights, but a full set of engineering assets, including inference deployment, datasets and evaluation, toolchains, workflows, and edge-to-cloud coordination. The parallel emergence of video generation tools, 3D components, distillation datasets, and agent frameworks pointed to something larger than isolated breakthroughs—it pointed to reusable system-level capabilities.

Models in the 0.5B–30B range were easier to run locally, fine-tune, and integrate into business systems and agent workflows. Leading players often used large MoE models in the 100B–700B range as capability ceilings or "teacher models," then distilled those capabilities down into many smaller models. This created a clear structure: a few very large models at the top, and many practical models underneath.

After R1, Apache 2.0 became close to the default choice for open models from the Chinese community. More permissive licenses lowered the friction around using, modifying, and deploying models in production. In 2025, model releases increasingly aligned with inference frameworks, quantization formats, serving engines, and edge runtimes. A prominent goal was no longer just to make weights downloadable, but to ensure that models could run directly on target domestic hardware—and run reliably and efficiently.

Training-side signals also began to appear. Ant Group's Ling models use optimized training on domestic AI chips to achieve near NVIDIA H800 performance. Baidu's open Qianfan-VL models were trained on a cluster of more than 5,000 Baidu Kunlun P800 accelerators. These disclosures showed that domestic computers were no longer limited to inference, but had started to enter key stages of the training pipeline.

On the serving and infrastructure side, engineering capabilities are being systematically open-sourced. Moonshot AI released its serving system: Mooncake. Baidu's FastDeploy 2.0 emphasized extreme quantization and cluster-level optimization. Alibaba's Qwen ecosystem pursued full-stack integration to minimize friction from development to production. These decisions signal a clear shift in competition from model performance to system design.