Qwen 3

Qwen 3, released by Alibaba in April 2025, is an open-source model family featuring both dense and Mixture of Experts (MoE) architectures. The flagship model, Qwen3-235B-A22B, has 235B total parameters with 22B active, while dense models range from 0.6B to 32B parameters and MoE variants include 30B-A3B and 235B-A22B configurations. This diverse family provides options for different deployment scenarios and resource constraints.

The models are trained on an impressive 36 trillion tokens spanning 119 languages, making Qwen 3 one of the most multilingual model families available. A key feature is hybrid reasoning, allowing the models to switch between thinking mode (slower, more deliberate reasoning) and non-thinking mode (faster direct responses) depending on task requirements.

Qwen 3 is competitive with DeepSeek R1, OpenAI's o1 and o3-mini, and Grok 3 across various benchmarks, demonstrating that strong performance can be achieved outside of American AI labs. The combination of open-source availability, multilingual capabilities, and flexible reasoning modes makes Qwen 3 particularly valuable for international deployments and applications requiring language diversity.