On April 7 (local time), Stanford University in the United States announced in the 'Artificial Intelligence (AI) Index 2025' that Alibaba has six notable AI models, ranking third in the world, following OpenAI and Google, which each have seven. However, it is the number one among corporations that disclose AI models as open source, surpassing Meta (four models). Starting from 2023, Alibaba has been releasing its large language model (LLM) Qwen series as open source, which means opening the 'source code' necessary for developing the program for anyone to use for free. This is in contrast to so-called 'closed AI models' that do not disclose how their AI models operate, like OpenAI and Google.
Alibaba's strength lies in building its own AI value chain. When Alibaba Cloud develops basic models, corporations utilize them in their own models or applications. In real life, Alibaba Group's B2C (business-to-consumer) marketplace, such as Taobao, operates a business using these AI solutions. Moreover, it is creating its own open-source ecosystem by establishing an open-source community.
The IT industry predicts that Alibaba's AI ecosystem will become more robust. In February, the company announced that it plans to invest over 380 billion yuan (approximately 74 trillion won) in cloud and AI hardware infrastructure over the next three years. This exceeds the total amount of Alibaba's investments over the past decade.
What does the future of the open-source AI model envisioned by Alibaba look like? I asked Yoon Yong-jun, who oversees the Alibaba Cloud business in Korea, about their open-source strategy, philosophy, and future roadmap. Here is the Q&A.
What does it mean that you recently released the AI model Qwen as open source?
"Over the past three months, we have been continuously releasing various AI models as open source. Notably, there is the large-scale language model (LLM) 'Qwen 2.5-Max', the strong model for complex questions 'QwQ', and the 'QVQ series', which is specialized in image understanding. These models not only enhance the completeness of AI technology but also demonstrate Alibaba's willingness to collaborate with more developers through open source."
How is the performance?
"The Qwen series has repeatedly ranked high on the performance comparison leaderboard (Open LLM leaderboard) operated by the world's largest open-source AI community, 'Hugging Face'. More than 100,000 derivative models have been developed based on it worldwide, surpassing the number of derivative models based on Meta's AI model 'LLaMA'. Currently, Qwen is one of the most widely used open-source AI models."
In March, you also released the open-source AI inference model 'QwQ-32B'.
"QwQ-32B is an inference-specialized AI model designed to solve complex problems step by step. It shows excellent performance in areas such as solving math problems and coding. This model incorporates the 'reinforcement learning' technique, allowing it to understand user requests more accurately and adjust responses according to human preferences. To enhance command execution capabilities, it combines a reward system with automatic validation features. Additionally, it performs a type of 'AI agent' role by using tools to find answers or adapt problem-solving methods based on feedback from the surrounding environment. QwQ-32B outperforms OpenAI's 'o1' and DeepSeek's 'R-1' in complex inference and analysis. We define this model as the world's first 'inference-specialized open-source AI' and continue to pursue a strategy of transparently sharing technology and developing together."
As it is open-source, there must be quite a few cases of utilization from outside.
"In fact, global collaboration related to open source is active. A research team led by Professor Fei-Fei Li at Stanford University developed an inference-specialized model 'S1' based on 'Qwen 2.5-32B', which reportedly cost less than $50 (about 70,000 won) to develop. UC Berkeley recreated DeepSeek's high-performance model at low cost using the 'Qwen 2.5' model and developed 'TinyZero'. The AI public research institute 'Allen Institute for AI' created the latest optical character recognition (OCR) model based on 'Qwen 2.5-VL-7B'. Qwen is playing a central role in leading a global ecosystem based on collaborative research, beyond just being open source."
Many corporations develop AI models in a closed manner; why do you adhere to open source?
"It relates to Alibaba's philosophy. Since our founding, we have supported the growth of small businesses and innovators under the mission of 'helping the growth of small companies'. We believe that AI technology should also be freely accessible to everyone. Thanks to this corporate culture, we can maintain a consistent open-source strategy. We see open source as a means to democratize technology and accelerate innovation. Going forward, we aim not only to disclose technology simply but also to actively realize the democratization of technology through open source."
What open-source strategy are you formulating?
"Our long-term strategy is to operate all processes, from AI model development to community contribution, transparently and scalable. This is the background for creating 'ModelScope', the largest open-source AI community platform in China, fostering an ecosystem where anyone can experiment with and utilize various technologies. Essentially, we are building a global collaborative ecosystem based on open source."
How do you plan to advance the open-source ecosystem in the AI field?
"There are four aspects. First, the democratization of AI. We aim to lower the barriers to access to the latest technologies so that under-resourced groups can also utilize AI. Second, the acceleration of scientific research. It will facilitate experimentation and reproducibility, thus advancing scientific progress. Third, transparency and trust. By disclosing the source code, we seek to solve the 'black box problem' of closed models that makes it difficult for users to understand the internal workings or decision-making processes of AI models. Fourth, the promotion of innovation and collaboration. We will help the global community improve existing models and create new ones."
As AI technology becomes more sophisticated, will the importance of open source grow?
"Absolutely. First of all, we believe that as AI models become more powerful, this technology should not be monopolized by a few. We believe open source is a key strategy to lead the next phase of AI. Through open source, researchers, startups, and small and medium-sized enterprises around the world should be able to participate in technological advancement and share its benefits. Creating an environment where everyone can grow and compete through technology is our targeted direction, which is in line with the intrinsic value of open source."
What is the open-source roadmap you are planning?
"We plan to continue expanding investments in basic research and open innovation. In the short term, we will enhance the performance of large basic models and make them easy for industry-specific customers to utilize. We will also expand the open-source ecosystem and refine the collaboration structure with the global community to strengthen the foundation for developers and researchers worldwide to advance technology together."
Chinese corporations are leading in open-source AI, coming second to the U.S. in the number of open-source AI models.
Recently, in the Chinese AI market, open source has established itself as a 'trend'. It requires less development expense and offers better accessibility than closed AI models, making it easier to secure market leadership. Not only Alibaba's Qwen series but also DeepSeek's LLM 'V3' and 'R-1', which gained attention earlier this year, are all released as open source. Recently, ByteDance, the parent company of TikTok, unveiled the video generation AI model 'OKONG' in collaboration with Hong Kong University as open-source. As a result, China has been found to hold the second-largest number of high-performance open-source AI models globally, after the U.S. According to Totus Media from the U.K., among the 200 high-performance open-source AI models released last year, the U.S. accounted for 65 and China 41. Notably, in regions with poor AI development conditions like Southeast Asia and Africa, China's open-source AI models are actively being adopted. This is because anyone can develop AI easily and cost-effectively. Thus, it is said that there is a national-level atmosphere in China that encourages the open source of AI technology.