The 2026 CloudCon was held in Hangzhou on September 22. Alibaba Group announced multiple advances in the field of large models: Qwen4, based on the next-generation architecture, has been put into training, and subsequent models such as Qwen4.5 and Qwen5 are planned to expand their parameter scale to 500 billion to 1 trillion; Recursive Self-Improvement (RSI) for large models has entered the stages of model training, inference, and chip collaboration. The multimodal model matrix has been fully upgraded, and the next-generation video generation model will be released in November.
Liu Dayiheng, head of the Qwen LLM project, stated that Scaling is an important path towards ASI. Alibaba revealed that after the release of Qwen3.8-Max, it ranked among the top in evaluations such as Artificial Analysis Agentic agents and CodeArena front-end programming. More notably, Qwen3.8-Max has been able to independently build training processes, construct data, design experiments, and identify defects, continuously iterating for over one month without human involvement, completing 33 effective iterations, which improved the Artificial Analysis score by 12.5%. In terms of reasoning optimization and chip model collaboration, RSI has also been implemented.
In terms of architecture, Qwen3.8-Flash reduces training costs by nearly 90% through innovations such as attention mechanisms, and improves inference efficiency with fewer parameters activated. In the multimodal aspect, Qwen3.8-Omni-Flash processes video, audio, images, and text uniformly; Wan3.0 supports generating a 30-second video at once and structured input, and the next-generation video model will be released in November. The Qwen-Audio-3.1 series covers ASR, TTS, and Realtime, while Qwen3.8-LiveTranslate reduces the average delay per character to 2.3 seconds. Qwen-Image-3.1, Qwen-Image-2.1, HappyShrimp 1.1, and HappyOyster 2.0-Preview are also being advanced simultaneously.
In the open-source ecosystem, within one month, the download count of Qwen3.8-related models has exceeded 56 million times, with over 1,900 derivative models. As of now, Alibaba has open-sourced more than 460 Qwen models, with a total download count exceeding 3 billion times and over 300,000 derivative models. Alibaba stated that companies such as Perplexity, Airbnb, Pinterest, and Reuters have used Qwen for applications including Agents, AI assistants, or self-developed models.
