Alibaba Launches Qwen3 AI Models to Rival Google and OpenAI
Alibaba introduces Qwen3, a new family of hybrid AI reasoning models that challenge top Western AI systems. The open-weight models range from 0.6 billion to 235 billion parameters and support 119 languages.
Alibaba releases Qwen3, a new family of AI models that matches and sometimes outperforms the best offerings from Google and OpenAI. The models range in size from 0.6 billion to 235 billion parameters and are available for download under an open license on platforms like Hugging Face and GitHub. This launch increases the competitive pressure on American AI labs and highlights the growing influence of Chinese-developed AI technologies.
The Qwen3 models function as hybrid systems that seamlessly switch between fast, non-thinking responses and slower, deliberate reasoning for complex problems. This self-fact-checking reasoning capability resembles the technology found in OpenAI's o3 model. Additionally, some models utilize a mixture of experts (MoE) architecture, which delegates subtasks to specialized smaller models to improve computational efficiency.
Alibaba trains these models on a massive dataset of over 36 trillion tokens using a diverse mix of textbooks, code snippets, and AI-generated data. The entire model family supports 119 languages, making it a highly versatile tool for global developers. By allowing users to easily configure task-specific thinking budgets, Alibaba delivers significant capability upgrades over its previous Qwen2 generation.