Alibaba unveils "Qwen3" — a hybrid reasoning model that switches between thinking modes
On April 29, 2025, Alibaba announced its next-generation model series, “Qwen3.” It is Alibaba’s first hybrid reasoning model, seamlessly switching within a single model between a “thinking mode” that works carefully through complex problems and a “non-thinking mode” that answers simple questions instantly. It was positioned as setting a new bar for open-source AI.
Details
- Lineup: Offered 8 models in total — 6 dense models (0.6B, 1.7B, 4B, 8B, 14B, 32B) and 2 MoE models (30B total parameters with 3B active, and 235B total parameters with 22B active)
- Hybrid reasoning: Qwen’s first design allowing the model to switch internally between a “thinking mode” for complex tasks and a “non-thinking mode” for fast responses
- Training scale: Trained on roughly 36 trillion tokens, twice the training data volume of the previous generation, Qwen2.5
- Multilingual support: Supports 119 languages and dialects, delivering strong performance in translation and multilingual instruction following
- License: Released all 8 models with open weights under the Apache 2.0 license, downloadable from Hugging Face, GitHub, and ModelScope
What happened next
Qwen3 was highly regarded as the culmination of Alibaba’s open-source strategy. But heading into 2026, Alibaba rapidly refreshed its top-tier models through Qwen3.5, Qwen3.6, and Qwen3.7, while gradually shifting its flagship-class models toward a closed, API-only strategy.