Shi Yuxiang is eager to find out what the latest artificial intelligence (AI) technology has to offer in video-making. But with OpenAI’s long-awaited Sora video generation tool launching only earlier this month, the 34-year-old entertainment industry professional from Beijing has been experimenting with a wide range of Chinese alternatives.
“Whenever there’s a new product release, I’ll give it a try,” Shi said. If a certain tool impressed him, he would pay for a subscription.
Shi is among the many tech-savvy Chinese users spoiled by a bevy of home-grown generative AI (GenAI) services, as tech giants and cash-flush start-ups battle for customers in a fast-growing market. As of November, regulators had approved 252 GenAI services for public release in the country.
Do you have questions about the biggest topics and trends from around the world? Get the answers with SCMP Knowledge, our new platform of curated content with explainers, FAQs, analyses and infographics brought to you by our award-winning team.
Chinese companies have been rushing to fill the void left by world-leading AI players from Microsoft-backed OpenAI to Google, whose GenAI services remain officially unavailable to the world’s largest internet population.
While mainland businesses initially fell behind their Western peers in the AI arms race triggered by OpenAI’s release of ChatGPT late in 2022, Chinese firms have moved up quickly this year.
An AI robot at a consumer product expo in Haikou, capital of southern Hainan province. Photo: Xinhua alt=An AI robot at a consumer product expo in Haikou, capital of southern Hainan province. Photo: Xinhua>
When OpenAI teased Sora early in February and provided limited access to a group of testers, it looked like China’s AI players, already hindered by escalating US chip curbs, were lagging behind.
The arrival of Sora was like a “barrel of cold water poured over China’s head”, said Zhou Hongyi, the founder of Chinese internet security firm 360 Security Technology.
But mainland companies scrambled within months to come up with their own Sora competitors.
Kuaishou Technologies, the main local rival of TikTok maker ByteDance and one of the first companies to release a GenAI video tool in China, opened its Kling service for a limited trial in June and expanded the test to global users in July.
The company also collaborated with nine Chinese film directors – including Jia Zhangke, a Golden Lion winner at the 2006 Venice Film Festival – to create short films using Kling.
In the months that followed, a group of local companies including state-backed Zhipu AI, Beijing-based Shengshu AI, ByteDance and social media and video game powerhouse Tencent Holdings introduced similar tools, touting improved video quality, more lifelike images and longer video lengths.
Analysts attribute the rapid development of AI products in China to a combination of factors.
Under President Xi Jinping, Beijing has made AI a national priority, driving substantial public and private investments in the field, according to Ray Wang, a Washington-based independent analyst focusing on US-China tech and economic relations and strategies.
Along with a robust talent pool in science, technology, engineering and mathematics, “these two factors have nurtured both AI start-ups and big tech companies, enabling them to advance AI development”, Wang said.
Apart from video tools, Chinese firms also launched a flurry of reasoning models that they claimed could match or even surpass OpenAI’s latest products in some areas.
After OpenAI released in mid-September a preview of its o1 reasoning model designed to “think and reflect” before giving out responses, Chinese firms introduced a quick succession of reasoning models, such as InternThinker from Shanghai AI Lab and Skywork o1 from Kunlun Tech, owner of the Opera browser.
The latest offering came from Alibaba Group Holding’s Qwen team, which unveiled its QwQ open-source visual reasoning model on Wednesday, calling it a “holiday gift”. The model is “closing the gap” with OpenAI’s o1 model, the team said. Alibaba owns the South China Morning Post.
An exhibit at the World Artificial Intelligence Conference in Shanghai. Photo: AFP alt=An exhibit at the World Artificial Intelligence Conference in Shanghai. Photo: AFP>
Earlier this month, ByteDance and Moonshot AI released their respective reasoning models equipped with visual perception capabilities, allowing them to “see and think”.
In November, DeepSeek debuted its R1 model, saying it outperformed the preview version of OpenAI’s o1 in half of six benchmarks tested by the start-up encompassing maths, programming and scientific exploration.
The swift pace of AI model releases highlights the fierce competition among Chinese companies in the area, according to Adina Yakefu, an AI researcher at the US-based machine learning community Hugging Face.
China has a massive market with a wide variety of application scenarios, offering a strong foundation for advancing AI, she added.
Chinese AI models are also starting to make waves outside their home market.
Open-source models from Hangzhou-based Deepseek and Alibaba’s Qwen team, accessible to the global AI community on platforms such as Hugging Face, are gaining traction, according to James Wong, general partner at San Francisco-based Creative Ventures.
On Hugging Face’s trending model section, which lists the most popular models on the platform over the past seven days, products built by or modified from those developed by Chinese companies made up half of the top 10 last week.
The platform’s most downloaded model this year, accounting for over a quarter of the total, was Alibaba’s Qwen 2.5.
By adopting an open-source approach, Chinese firms have taken “a smart direction”, Wong said. Since these companies did not start in a leading position like OpenAI or Google, it made sense for them to make some of their products more accessible to entice users.
“You need some inducements for companies to use your product,” he said.
As of mid-September, more than 50,000 open-source model variants were built on Qwen, making it the world’s second-largest AI model ecosystem after Meta Platforms’ Llama, Alibaba Cloud said at the time.
A team of researchers from Meta and Stanford University used Qwen 2.5 “at varying scales to serve as the backbone” for its new multimodal model with video generation capability called Apollo, according to a paper published earlier this month.
Apollo had outstanding performance across multiple model sizes, frequently outperforming models two to three times their sizes, the researchers said.
“China will start to lead the AI race as a consequence of leading the open-source AI race,” Hugging Face co-founder and chief executive Clem Delangue said in a recent post published to his LinkedIn account.
For now, the world’s top AI models are still mostly closed-source. These include Google’s latest Gemini 2.0 and OpenAI’s GPT-4o and o1 series, according to Chatbot Arena, an AI benchmarking platform developed by UC Berkeley researchers.
However, advanced open-source models from DeepSeek and Qwen were quickly closing in on the world’s top five closed-source models in average performance, the SuperClue benchmark test found in a report published in November.
Chinese companies believe that open-sourcing their models can help them build a stronger ecosystem, according to DeepSeek CEO Liang Wenfeng.
“We won’t go closed-source,” he said in an interview with Chinese tech news outlet 36Kr. “We believe that having a strong technological ecosystem is more important.”
Despite the rapid progress made by China, analysts warn that US trade restrictions on advanced chips – such as Nvidia’s premium graphics processors, which have become the go-to options for training and running AI models – could be the Achilles’ heel for Beijing’s AI ambitions.
At the moment, Chinese tech firms are still able to train their models on Nvidia chips stockpiled before Washington’s export curbs kicked in, such as the since-restricted A100 chips, according to analyst Wang.
In addition to those supplies, some Chinese entities also acquired chips from other countries and through smuggling, loopholes that Washington is looking to plug by capping graphics processor shipments to specific locations and implementing a worldwide licensing system, according to a Post report earlier this month.
However, Chinese companies “will ultimately need new hardware in the next few years, and they will thus face the ‘AI hardware bottleneck’ as they are not able to acquire more advanced AI chips than the ones they currently possess”, Wang said.
A lack of sophisticated AI chips could eventually “widen the performance gap between China and the US in AI development”, he said.
This article originally appeared in the South China Morning Post (SCMP), the most authoritative voice reporting on China and Asia for more than a century. For more SCMP stories, please explore the SCMP app or visit the SCMP’s Facebook and Twitter pages. Copyright © 2024 South China Morning Post Publishers Ltd. All rights reserved.
Copyright (c) 2024. South China Morning Post Publishers Ltd. All rights reserved.
