AI Trends

Chinese Large Models Lead the Global Open-Source Ecosystem

People’s DailyUpdated Nov 21 06:06

AI summary

Chinese open-source large models are expanding their deployment in industrial manufacturing, application development, and overseas markets through open weights, architectural innovation, and compatibility with domestically produced computing infrastructure. From November 9 to 15, Chinese models recorded 119.1 trillion tokens in weekly usage on OpenRouter, ranking first globally for the 29th consecutive week; cumulative global downloads of China’s open-source AI large models surpassed 10 billion. Moonshot AI released the complete model weights for Kimi K3.1 Pro, while Tencent Hunyuan released Hy4, and numerous companies are also exploring advances in model architecture, inference costs, and application depth.

Why read this

AI score 60

The article reviews open-source trends among Chinese models from three perspectives—open weights, architectural innovation, and industrial applications—and presents their global usage and deployment.

Topics

Tags

Body · AI translation

Recently, Beijing Moonshot AI Technology Co., Ltd. released the complete model weights for Kimi K3.1 Pro, the world's largest open-source model by parameter count, sparking widespread discussion in the technology community. Tencent Hunyuan also released Hy4, its next-generation large language model, with improvements in instruction following, long-document comprehension, long-horizon tasks, deep research, and other capabilities.

At the recently concluded 2026 World Internet Conference Wuzhen Summit, “open source” was also a key theme. Since the beginning of this year, numerous Chinese technology companies have intensively launched open-source model products, continually achieving breakthroughs in parameter scale and technological depth. What is open source? How should the trend of open-sourcing models be viewed? What new development trends are emerging among Chinese open-source models?

Open-source efforts accelerate, leading in downloads and derivative models

Unlike the literal meaning of “making source code publicly available” in software, open-sourcing a model more often means providing “open weights.”

Put simply, a model is like a brain containing a vast amount of knowledge, while its “weights” are the hundreds of billions of parameter values within that brain. These parameters record all the linguistic patterns, logical reasoning abilities, and knowledge that the model learned during training.

“Open weights” means that users can download a model, run inference on their own computers or servers, and even fine-tune it for specific tasks. For example, companies and developers can obtain an extremely capable foundation model without spending heavily to train one from scratch. It can also be deployed on local servers or private clouds, which is crucial for users with particularly stringent data-privacy requirements. An increasingly diverse range of open-source models also enables developing countries to access cutting-edge capabilities at low cost.

Open source, however, does not mean relinquishing commercial returns. Through cloud inference, agent platforms, model fine-tuning, and other services, companies can still build sustainable revenue models.

China’s average daily token usage has now exceeded 500 trillion. Domestic online shopping platforms have launched token top-up centers, and tokens even became hot commodities in livestreaming rooms during the “Double 11” shopping festival. “The measurable, priceable, and tradable attributes of tokens are gradually being uncovered, and token usage is becoming an important indicator for gauging artificial intelligence (AI) activity, the depth of industrial penetration, and the level of intelligent development,” said Wang Jingtao, deputy director of the Cyberspace Administration of China.

Chinese open-source models have also performed exceptionally well on highly market-oriented model aggregation platforms. From November 9 to 15, Chinese models recorded weekly usage of 119.1 trillion tokens on OpenRouter, a global open-source model aggregation platform, retaining the top global position for the 29th consecutive week. Data from the Ministry of Industry and Information Technology show that Chinese open-source AI large models have surpassed 10 billion cumulative downloads worldwide. “A group of world-leading open-source models continues to emerge, leading in downloads and the number of derivative models and driving the global open-source AI ecosystem,” said Huang Qi, secretary of the Communist Party of China Committee at the Institute of Scientific and Technical Information of China.

Growing competitiveness as model architecture innovation accelerates and foundational technologies achieve continued breakthroughs

The growing competitiveness of Chinese open-source models stems not only from their broad range of applications, but also from faster innovation in model architectures and continued breakthroughs in foundational technologies.

Early last year, the DeepSeek series of models released by DeepSeek achieved breakthroughs not only in performance but also through innovations in the mixture-of-experts model architecture, substantially reducing inference costs and computing-power requirements. This architecture later became a shared area of exploration across the global AI sector.

Kimi K3, released this year, introduced a proprietary hybrid linear attention mechanism and other technologies, making “ultra-long text more economical to process and ultra-large models more stable to train.” “We reconstructed key components of the architecture from the level of its underlying logic, opening up entirely new possibilities for the evolution of model training,” said Yang Zhilin, founder and CEO of Moonshot AI’s Kimi. Chinese open-source models are becoming benchmarks in evaluations by global model developers and important research samples for overseas institutions.

The improving compatibility between Chinese open-source models and domestically produced computing power also merits attention. In June this year, Meituan released a trillion-parameter large model whose entire training and inference processes were completed using domestically produced computing power. In November, Meituan released its next-generation large model, LongCat—2.5, advancing large models on domestic computing infrastructure from merely being “able to run” to “running fast and efficiently.” “We have gradually overcome challenges including deterministic computation on neural processing units and improvements in computing-power utilization, validating the feasibility of large-scale training on domestic hardware,” said Cai Xunliang, head of Meituan’s LongCat large model.

To promote the practical deployment of AI, domestic chipmakers are also continuously optimizing their product designs. “The growing capabilities of Chinese open-source models are accelerating demand for agents, programming, video generation, and other applications. This will generate large volumes of high-frequency calls involving long contexts and multiple rounds of tool use, as well as demand for more efficient concurrent multimodal processing. The larger the usage volume, the more sensitive users become to response speed and cost,” said Xu Bing, chairman of Si-Way, a Chinese manufacturer of inference GPUs (graphics processing units).

Deeper applications bring the benefits of intelligent innovation to the world

As technology develops, Chinese open-source models are finding their way into more application scenarios.

Late at night on a BOE production line, display panels proceed methodically through each stage of production. An AI system detects a microscopic defect in real time, immediately generates a repair plan, and coordinates with equipment to complete the repair—from detecting and analyzing the problem to making a judgment and carrying out the response, it creates a fully intelligent, end-to-end closed loop for quality management.

“Using a Chinese open-source large model as our foundation, we independently developed BOE’s Blue Whale large model for displays. It integrates industrial manufacturing expertise and production-line data, covers scenarios including R&D and design, manufacturing, and quality management, and is helping deploy numerous AI applications at BOE factories across China,” said Jiang Xingqun, senior vice president and co-chief technology officer of BOE.

When independent developer Wu Ruimeng received a request from a Taobao shop owner to “wrap a car using anime-style images,” he immediately thought of the “AI virtual try-on” scenario. Using the no-code application development platform Miaoda, he designed a demo in just 25 minutes. This “+AI” concept ultimately sold for 8,000 yuan. “By describing requirements in natural language and having a model create a customized application, AI can help make up for gaps in my capabilities,” Wu said.

Chinese open-source models are likewise transforming how overseas users work and the cost structures of overseas companies.

Late last year, a multilingual large model called “Sea-Lion V4” was released in Singapore. Developed on the foundation of Alibaba’s open-source Qwen model, it focuses on Southeast Asia’s complex linguistic environment and has been well received by local users. Building on Qwen’s advantage of supporting 119 languages, “Sea-Lion V4” was pretrained on 36 trillion tokens. In the view of Liu Sili, senior director at AI Singapore, collaboration with Chinese companies is an important milestone in making AI broadly accessible and more representative of Southeast Asia.

“The possibility of making AI broadly accessible must be built on a solid open-source foundation,” said Zhang Yunquan, a researcher at the Institute of Computing Technology, Chinese Academy of Sciences. From technology and rules to public goods, China is turning the vision of broad accessibility into concrete action.

Source

People’s Daily