Skip to content
Tech

Chinese open-source large models lead global ecosystem with over 10 billion downloads

China News Editorial team · AI-written, reviewed under our editorial standards · · Reading time 10min read · Views 0 ·
Key — Chinese open-source large models lead global ecosystem with over 10 billion downloads.

The global cumulative downloads of Chinese open-source large models exceeded 10 billion. The Ministry of Industry and Information Technology reported that the figure for downloads was reported by media reports. This figure is reported by media reports.

Open models and token calls

Open weights allow users to download models, run inference on their own hardware, and fine-tune them for specific tasks. The OpenRouter platform reported that Chinese models received 1.191 trillion weekly token calls between November 9 and 15. This figure was reported by media reports.

On the OpenRouter platform, Chinese models ranked first for 29 consecutive weeks. The Ministry of Industry and Information Technology reported that the global cumulative downloads of Chinese open-source large models exceeded 10 billion.

Model releases and architecture

Beijing Moonshot AI Technology Co., Ltd. released the full model weights for Kimi K3.1 Pro. This model is the largest open-source model by parameter scale. The 2026 World Internet Conference in Wuzhen highlighted open source as a key theme.

Tencent Hunyuan released the next-generation large language model Hy4. Hy4 improves instruction following, long-text understanding, long-range tasks, and deep research.

DeepSeek released the DeepSeek series early last year. The series innovated the Mixture-of-Experts architecture to reduce inference costs. Kimi K3 introduced a self-developed mixed linear attention mechanism to make long-text processing more efficient and large-model training more stable.

Meituan released a trillion-parameter model trained and inferred entirely using domestic computing power in June. The company followed this with LongCat-2.5 in November.

Applications and context

The Beijing Oriental Display factory used a domestic open-source large model as a base to develop the Oriental Display Blue Whale Display large model for quality management. A developer used the no-code platform MiaoDa to create a demo in 25 minutes for an 8,000 yuan order. The HaiShi V4 multilingual model, based on Alibaba Qwen open-source model, was released in Singapore last year and was pre-trained on 36 trillion tokens.

Wang Jingtao, Deputy Director of the Cyberspace Administration of China, said token calls are becoming a key indicator of AI activity, industry penetration, and intelligent development. Huang Qi, Party Secretary of the China Institute of Science and Technology Information, said global leading open-source models are emerging, leading in downloads and derivative models.

Yang Zhilin, Founder and CEO of Moonshot AI, said the architecture of Kimi K3 has been reconfigured to find new evolutionary spaces for model training. Cai Xunliang, Head of Meituan LongCat model, said Meituan has overcome challenges in deterministic computation and utilization rate to verify large-scale domestic training.

Xu Bing, Chairman of Xiwang, said increased open-source model capability will drive demand for long-context, multi-turn, and multi-modal processing. Jiang Xingqun, Senior Vice President and Co-CTO of Beijing Oriental Display, said the company developed the Oriental Display Blue Whale Display large model using a domestic open-source base.

Wu Ruimeng, a developer, said AI can fill capability gaps by creating customized applications from natural language descriptions. Liu Sili, Senior Director of the Singapore AI Programme, said cooperation with Chinese enterprises is an important sign for promoting AI inclusivity in Southeast Asia.

Glossary

  • Open models — Open weights allow users to download models, run inference on their own hardware, and fine-tune them for specific tasks. The OpenRouter platform reported that Chinese models received 1.191 trillion weekly token calls between November 9 and 15. This figure was reported by media reports.
  • OpenRouter — The OpenRouter platform reported that Chinese models received 1.191 trillion weekly token calls between November 9 and 15. This figure was reported by media reports.
  • Tencent Hunyuan — Tencent Hunyuan released the next-generation large language model Hy4. Hy4 improves instruction following, long-text understanding, long-range tasks, and deep research.
  • Hy4 — Tencent Hunyuan released the next-generation large language model Hy4. Hy4 improves instruction following, long-text understanding, long-range tasks, and deep research.
  • DeepSeek — DeepSeek released the DeepSeek series early last year. The series innovated the Mixture-of-Experts architecture to reduce inference costs.

This article was written with AI assistance from multiple news reports and is reviewed by our editors under our editorial principles. Editorial Standards · Tips & Contact

How did you like this post?

Comments 0

Be the first to comment

Contact us

← China News Home
China News Get new posts by emailSubscribe to receive new content via email. Unsubscribe anytime.
Was this helpful?Share it with friends & social