At the 2026 Apsara Conference, held from September 22 to 24 in Hangzhou, Liu Dayiheng, the project lead for the Qwen LLM under Alibaba's ATH Business Group's Token Foundry, took the stage to deliver a keynote address.

Liu highlighted that this year marks the formal entry of large language models into their third phase: the era of intelligent agents. He explained that models are no longer just assistants passively awaiting prompts; instead, they have evolved into agents capable of autonomous planning, tool invocation, and closed-loop iteration to independently complete tasks.

The consumption pattern of AI has also shifted from single interactions to the execution of complete tasks. This transformation has directly fueled an explosive, exponential surge in global demand for tokens.

Over the past few years, the team has consistently open-sourced advanced models across all modalities and sizes to developers worldwide. To date, the Qwen series has surpassed 3 billion global downloads, with over 300,000 derivative models, ranking first globally. Officially, more than 460 models have been open-sourced cumulatively.

Liu also noted that just recently, the Qwen 3.8 27B model, released only two months ago, has outperformed notable models such as DeepSeek and Llama, becoming the most-liked model on the HuggingFace platform. Qwen is building a truly global developer ecosystem.

"We have always believed that the value of AI should not belong solely to a select few enterprises but should become a foundational capability accessible, usable, and creatable by users worldwide. In the future, we will persist in open-source and open development, enabling more people to afford AI, utilize AI, and participate in creating AI, thereby achieving widespread AI accessibility," he stated.