Amid escalating competition with the United States, Alibaba and Zhipu AI have unveiled new and advanced open-source models, highlighting China’s significant progress in this vital field.
Article contents:
- Introduction: China accelerates the pace of innovation
- Alibaba’s Wan2.2: Video based on Mixture of Experts architecture
- Zhipu’s GLM-4.5: An advanced language model
- The open-source trend
- China’s progress in numbers
- Conclusion: Open source as a global battlefield
Introduction: China accelerates the pace of innovation
Alibaba Group Holding and Zhipu AI have launched new open-source models as competition between China and the United States intensifies in the field of artificial intelligence. These releases point to a broader industry trend toward open-source development, which enables developers to improve efficiency and achieve wider adoption of their products.
Alibaba’s Wan2.2: Video based on Mixture of Experts architecture
On Tuesday, Alibaba released Wan2.2, which it claimed is “the industry’s first open-source large video generation models incorporating the Mixture of Experts (MoE) architecture.” MoE is a machine learning approach that splits an AI model into separate sub-networks, or “experts”—each focusing on a subset of input data—to perform a task jointly. It enables models to be pre-trained with significantly less computing power.
The Wan2.2 series features text-to-video models, image-to-video models, and a hybrid model supporting both.
Zhipu’s GLM-4.5: An advanced language model
Separately, Zhipu, one of China’s four “AI tigers,” launched a new generation of its GLM series on Monday. The new models, which include GLM-4.5 with 355 billion parameters and a streamlined version GLM-4.5-Air with 106 billion parameters, are built on a fully self-developed architecture, according to the company.
Zhipu said on Monday that GLM-4.5 was “China’s most advanced open-source MoE model,” ranking third globally and first among domestic and open-source models based on average scores across “12 benchmark datasets.”
The open-source trend
Zhipu said its flagship open-source models have accumulated more than 40 million downloads worldwide since 2020. Similarly, Alibaba has embraced an open-source approach for its Qwen family of AI models starting in August 2023. Since then, more than 200 versions have achieved over 300 million global downloads.
China’s progress in numbers
China has made significant progress in AI models, highlighted by the global popularity of DeepSeek’s open-source reasoning model R1 launched in January. China is now home to 1,509 AI models, accounting for more than 40 percent of the global total, according to data presented at the World Artificial Intelligence Conference, which concluded in Shanghai on Monday.
Conclusion: Open source as a global battlefield
The new releases from Alibaba and Zhipu demonstrate that China is not only striving to catch up in artificial intelligence, but also aiming to lead the open-source movement. By sharing their models with the world, these companies hope to accelerate innovation and build developer communities around their technology, making open source a key battlefield in the global competition for AI dominance.
Frequently Asked Questions
What is the Mixture of Experts (MoE) architecture?
It is an AI technique that allows larger and more capable models to be built more efficiently. Instead of having one massive model, you have many smaller “expert” models, and only the appropriate experts are activated for a given task, saving computing power.
Who are China’s four “AI tigers”?
It is a term used to describe four of China’s most prominent artificial intelligence startups: Baichuan, Zhipu AI, Moonshot AI, and MiniMax.
Why is open source so important in artificial intelligence?
Open source allows the global community of researchers and developers to inspect, improve upon, and build on code, accelerating the pace of innovation. It also promotes transparency and helps discover bugs and biases faster.
