Overview
A report in People’s Daily on November 22 quotes an official of the Chinese Institute of Science and Technology Information saying that a batch of globally leading open-source models has emerged, with download counts and the number of derivative models at the forefront — a sign that China is shaping the global open-source AI ecosystem. The report attributes the gains to two threads: faster architectural innovation and tighter integration with domestic computing power.
On architecture, DeepSeek’s series applied innovations to the mixture-of-experts design that substantially reduced inference cost and compute demand, a direction later explored worldwide. Kimi K3 introduced a self-developed hybrid linear attention mechanism aimed at making ultra-long contexts cheaper to process and very large models more stable to train. On compute, after completing full-pipeline training of a trillion-parameter model on domestic hardware in the first half of the year, Meituan released LongCat-2.5 in November, pushing domestic-compute models from merely runnable toward fast and economical.
Background & Interpretation
1. Technical background and evolution
Competition among open models has shifted from raw parameter count toward architectural efficiency, inference cost and deployability. Mixture-of-experts designs separate training and inference economics through sparse activation, while hybrid linear attention targets the computational complexity of long contexts. Both resonate globally because they address the industry’s two hardest constraints: cost per token and hardware availability. A large developer ecosystem in turn accelerates iteration, reinforcing the cycle.
2. Core drivers and mechanisms
Model-compute co-design is the central mechanism. Inference GPU vendors note that stronger open models will accelerate demand for agents, coding and video generation, generating heavy long-context loads, frequent multi-round tool calls and multimodal concurrency. The larger the call volume, the more sensitive users become to latency and cost. That pushes chip designers to account for long-context and tool-call patterns at product-definition stage rather than chase peak specifications. Industrial deployment offers a proving ground: BOE built a display-manufacturing model on a domestic open foundation that detects microscopic defects in real time and generates repair instructions, closing the loop from detection to execution.
3. Market structure and value-chain impact
The ecosystem data matter as well. CCTV reported the same day that value added by the computer, communications and other electronic equipment manufacturing sector grew 15.8% year on year in the first three quarters, with profits doubling to 985.2 billion yuan. Adoption of generative AI products has passed 50%, sales of smart glasses on major platforms rose 110% and sales of embodied robots grew 22.6%. Chinese open models are becoming a benchmark for overseas researchers and a reference point for global vendors, while the openHarmony ecosystem has passed 1.35 billion connected devices — evidence of growing coordination between software foundations and applications.
Implications & Outlook
The open question is whether influence converts into a commercial loop: downloads and derivatives signal reach, not necessarily revenue. Key variables include actual delivery capacity of domestic inference chips, the pace of cost reduction in long-context and agent scenarios, and whether industrial users can accumulate reusable sector models. If hardware supply or cost advantages weaken, the ecosystem lead could be quickly contested.
Takeaways for Industry Participants
Application companies should treat open models as replaceable foundations rather than single dependencies, designing architectures that allow switching across both models and hardware. Chip and infrastructure vendors should define products around long contexts, tool calls and multimodal concurrency instead of competing on peak specifications. Industrial users will find it easier to validate returns by starting with tightly scoped loops such as quality inspection and document processing.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.