The global AI industry now shows parallel development of open and closed models. The degree of openness is not a simple binary choice — from fully open weights allowing arbitrary modification, to semi-open architectures keeping core parameters, to controlled API access, to fully closed black boxes. Competition between open and closed models has evolved from mere technical performance into a comprehensive contest of performance, ecosystem, security and greenness.
On inclusiveness, open models greatly lower the barriers to AI application, development and innovation, spreading capability to small and medium enterprises, research institutions, developers and traditional industries so they can build applications suited to local languages, industries and needs. By July 2026, China's open frontier models had again narrowed the gap with US closed frontier models and surpassed them on some core comprehensive metrics.
According to the Hugging Face 2026 summer observation report, China's Moonshot AI Kimi, Zhipu GLM and Alibaba Qwen series entered the global top seven; Kimi K3, with 2.78 trillion parameters, is the world's largest open-source model and leads on intelligence and coding indices. A joint MIT–Georgia Tech study on the OpenRouter platform found closed models cost on average 1.86 US dollars per million input-output tokens, while equivalently performing open models cost only 0.23 dollars.
Lightweight open models can run directly on high-performance consumer devices without remote data-center compute, further lowering edge-scenario thresholds. General frontier models are evolving toward a "general base engine + vertical specialist wings" hybrid architecture, where open models — customizble, privately deployable and lower-cost — hold the advantage in real scenarios.
Source: China Daily
