Overview
Qwen3, released by Alibaba’s Qwen team in April 2025 under Apache 2.0, spans dense models from 0.6B to 32B and mixture-of-experts models (30B-A3B and 235B-A22B). Each can switch between a step-by-step “thinking” mode and a fast mode.
Highlights
- Sizes from 0.6B to 235B, dense and MoE
- Hybrid thinking and non-thinking modes
- Support for over 100 languages and dialects
- Permissive Apache 2.0 license
Typical uses
Local assistants at every hardware level, reasoning tasks, coding and multilingual applications.
Things to consider
- Thinking mode produces long outputs that cost time and tokens.
- Native context is 32K–40K tokens; longer contexts need extension techniques.
- Download sizes refer to the default 4-bit quantized versions in the Ollama library; full-precision weights are larger.
Specifications are taken from the model card on Hugging Face, the Ollama library and the provider’s announcement. Benchmark results depend on evaluation settings; see how AI benchmarks work.