At the Apsara Conference in Hangzhou, hosted by Alibaba Cloud, China’s emerging AI startups highlighted their efforts in developing large language models, inspired by the recent launch of OpenAI’s o1 model. Backed by Microsoft, the o1 LLM is designed to tackle complex scientific, coding, and mathematical tasks. Industry leaders believe this advancement can accelerate AI innovation and open doors for new business opportunities.
Moonshot AI founder Kunal Zhilin emphasized the potential of the o1 model to transform industries, pointing to reinforcement learning and scalable model training as key factors. Zhilin noted that bigger models with more data typically yield better performance, a principle known as the scaling law. He added that reinforcement learning helps generate synthetic data, a growing necessity as organic data becomes scarce.
However, challenges remain. StepFun CEO Jiang Daxin pointed out that many Chinese startups struggle with limited access to high-performance computing, worsened by U.S. trade restrictions on advanced semiconductors. Only a few startups — including Moonshot AI, Baichuan AI, Zhipu AI, and MiniMax, collectively known as the “AI tigers” — have the resources to invest heavily in this field.
Alibaba Cloud also made major announcements at the event, including the release of its Qwen 2.5 model family. These models, ranging from 0.5 to 72 billion parameters and supporting 29 languages, include specialized versions for coding and mathematics, which have already seen over 40 million downloads. The company further showcased a new text-to-video generation feature in its Tongyi Wanxiang tool and introduced Qwen 2-VL, a vision-language model designed for mobile devices and robotics, capable of processing videos longer than 20 minutes.
These developments highlight a surge of AI innovation in China, fueled by both international breakthroughs and domestic tech leadership.
