Qwen3 Puts Reasoning on a Switch

Alibaba's Qwen team released Qwen3 in April 2025 as a full lineup under the Apache 2.0 license: dense models from 0.6B to 32B, plus two mixture-of-experts models, Qwen3-30B-A3B with 3B active parameters and the flagship Qwen3-235B-A22B with 22B active. It was pretrained on roughly 36 trillion tokens across 119 languages, about double Qwen2.5's data. The design choice that stands out is a single model with switchable thinking and non-thinking modes plus a configurable thinking budget, so you decide per request how much the model reasons before answering instead of running a separate reasoning model. Qwen reports the 235B flagship as competitive with DeepSeek-R1, OpenAI o1 and o3-mini, Grok-3, and Gemini 2.5 Pro, and says the small Qwen3-30B-A3B beats the older QwQ-32B while activating about a tenth of the parameters. These are the lab's own numbers, so treat them as a starting point for your own testing.

Why it matters

A permissive license across a wide size range makes Qwen3 an easy open default, and folding reasoning into a per-request switch is a practical pattern: you pay for thinking only when the task needs it, with one model to deploy instead of two.