Term
Qwen 3
Qwen 3 is Alibabas family of open models released in April 2025. It ranges from 0.6B to 235B parameters (dense and MoE models), is licensed under Apache 2.0 and unites a thinking and a non-thinking mode in a single model.
Qwen 3 — explained in detail
Qwen 3 is the third generation of the open Qwen models from the Chinese conglomerate Alibaba, released in April 2025. The models are under the permissive Apache 2.0 licence and are commercially usable without licensing fees as well as self-hostable (open-weight).
The family comprises eight variants from 0.6 billion to 235 billion parameters. There are dense models (0.6B, 1.7B, 4B, 8B, 14B, 32B) and Mixture-of-Experts models (30B-A3B and 235B-A22B). The flagship Qwen3-235B-A22B activates only around 22 billion of its 235 billion parameters per token. Depending on size, the context windows range from 32,000 to 128,000 tokens.
A distinctive feature is hybrid operation: Qwen 3 unites a thinking mode (step-by-step reasoning before the answer) and a non-thinking mode (direct answer) in a single model. Via the API, the thinking duration can be controlled granularly, enabling a deliberate balance between answer quality and compute cost.
Example / practical relevance
With its broad range of sizes, Qwen 3 covers very different use cases: the small models run on end devices or single GPUs, while the large MoE models deliver strong performance at moderate compute per token. The switchable thinking mode allows the same model to be used sometimes for fast answers and sometimes for demanding reasoning.
Distinction
As an open-weight family under the Apache 2.0 licence, Qwen 3 sets itself apart from proprietary models (GPT, Claude, Gemini). Among open families (Llama, Gemma, DeepSeek), Qwen 3 stands out through its wide span of model sizes and the dual-mode operation unified in a single model. Specialised offshoots such as Qwen3-Coder are separately geared towards coding.