Alibaba Unveils Qwen 3: A Smarter, Faster Multilingual AI Model

On 28 April 2025, Alibaba Group has unveiled Qwen 3, the latest upgrade to its large language model family, offering faster, smarter, and more globally accessible AI capabilities.

The models introduce hybrid reasoning which makes it one of the most versatile open-source models for tasks like translation, coding, and general conversation.

Key Points

• Qwen 3 introduces major improvements in reasoning and multilingual support, now working with 119 languages and dialects.
• Qwen3 offers six dense models and two Mixture-of-Experts (MoE), models, ranging with 600 million, 1.7 billion, 4 billion, 8 billion, 14 billion and 32 billion parameters.
• The new MoE models are designed for efficiency—Qwen3-30B uses just 3 billion active parameters at a time, and the 235B version uses 22 billion, making them faster than traditional dense models if there’s enough RAM.
• Qwen3 uses a dual-mode system that lets users switch between “thinking mode” for complex tasks like coding, and “non-thinking mode” for faster, casual conversations—helping balance speed and depth.
• Qwen 3 can handle complex tasks that involve tools or step-by-step reasoning, positioning it as a top-tier option for AI agents and applications.

Background

Alibaba unveiled Qwen3 just three months after the release of Qwen2.5-Max, underscoring the rapid pace of competition in the generative AI space.

According to the official blog, Qwen3’s base model was trained on 36 trillion tokens—double the 18 trillion used for Qwen2.5—marking a significant increase in data volume.

The training data included not only HTML content but also text extracted from PDFs using Qwen2.5-VL. To further enhance the model, Alibaba leveraged its earlier Qwen2.5-Math and Qwen2.5-Coder models to generate synthetic training data, reflecting a layered and efficient approach to model development.

News Gist

Alibaba launched Qwen 3, a powerful upgrade to its AI model family.

Supporting 119 languages and hybrid reasoning, it includes efficient Mixture-of-Experts models and a dual-mode system for speed or depth.

Trained on 36 trillion tokens, Qwen 3 marks a major leap in multilingual, tool-augmented AI.