Southeast Asia Information Port News (www.dnyxxg.com) On Chinese New Year's Eve, Alibaba open-sourced its new generation of large-scale model, Qwen3.5-Plus. This version boasts 397 billion parameters, with only 17 billion activated, reducing deployment memory usage by 60%, optimizing speed and cost while maintaining capabilities.
Qwen3.5-Plus's overall capabilities rival those of the Gemini3 Pro, setting new records for open-source models in multiple authoritative benchmark tests. Its API (Application Programming Interface) price is as low as 0.8 yuan per million tokens, only 1/18th of that of the Gemini3 Pro.
Notably, Qwen3.5 achieved a breakthrough in native multimodal computing by pre-training on mixed text and visual data. It performed exceptionally well in comprehensive benchmark evaluations across inference, programming, and agent capabilities, and achieved several best-in-class performance awards in authoritative assessments of visual understanding.
Before the Spring Festival holiday, Chinese tech companies had already begun a flurry of new product launches: Zhipu announced the launch and open-sourcing of its GLM-5 model, calling it "the best open-source model of the Agentic Engineering era"; iFlytek and MiniMax also released new models: the Spark X2 large-scale model trained with entirely domestic computing power, and the M2.5 model, respectively. The M2.5 model demonstrated higher decision-making maturity when handling complex tasks.
ByteDance's next-generation AI video generation model, Seedance 2.0, has also recently come into the spotlight, quickly attracting significant global attention with its multimodal creation methods and built-in camera movement effects.
This isn't the first time Chinese tech companies have bet on the Spring Festival season. Looking back at the 2025 Spring Festival holiday, Alibaba Cloud's Tongyi team released Qwen2.5-Max, a new version upgrade of its large-scale model, and DeepSeek saw a surge in new platform integrations, further boosting its popularity.
Hu Yanping, a specially appointed professor at Shanghai University of Finance and Economics, believes that the new models released this year are becoming increasingly "versatile and practical," featuring native multimodal and intelligent agent capabilities, balanced performance in code and multilingual support, and even usability in spatial intelligence and visual reasoning. They require only one-third the size of previous models while achieving the same "quality," yet decoding throughput has increased several times over. All of this simplifies deployment and usage, significantly improving efficiency.
Industry experts believe that the significance of Qwen 3.5 goes beyond simply "refreshing the rankings." In the past two years, the large-scale model industry has believed in "great effort yields miracles," pushing parameters from hundreds of billions to trillions. Performance has increased, but so has the cost. Deploying dedicated application clusters and inference requires massive computing power, making it unaffordable for small and medium-sized enterprises. The technology has become increasingly powerful, but truly usable has become increasingly distant. Qwen 3.5's approach is not about size, but about intelligence—who can use a relatively smaller model to achieve greater intelligence through technological innovation.
Integration is another major highlight, far exceeding simply "understanding a single image." Qwen 3.5 can perform pixel-level spatial localization and code-level fine-grained processing of images, understand the temporal evolution and causal relationships in videos up to 2 hours long, directly convert a hand-drawn sketch into executable front-end code, and even autonomously control mobile phones and computers as a visual intelligent agent to complete complex multi-step tasks across applications.
This progression from recognition to understanding, and from single-modal to cross-modal reasoning, lays a solid foundation for more natural and consistent multimodal generation and reasoning in the future.
Currently, there are over 400 open-source Qwen models, with over 200,000 derivative models and over 1 billion downloads, far exceeding Llama in the United States. Global companies and institutions, including Fei-Fei Li's team and Airbnb, are using Qwen models.
Ni Guangnan, an academician of the Chinese Academy of Engineering, recently stated that open source has become a powerful driving force for the development of global information technology, and its performance is strong in the emerging field of artificial intelligence led by large models. China has become the world's largest provider of open-source large models, with Chinese models such as Qwen, DeepSeek, and Kimi ranking highly on the AI model evaluation platform LMArena. He also mentioned that some open-source communities led by Chinese companies are thriving internationally, demonstrating great vitality in the AI era. Chinese companies are actively embracing the open-source philosophy and have become leaders in global open-source model innovation. (End)
