Qwen3-4B is a four-billion-parameter dense language model released by the Qwen team. Its model card describes multilingual support and the option to switch between thinking and non-thinking modes.
Qwen3-4B within the model family
The Qwen3 family includes both dense and mixture-of-experts models. Qwen3-4B is one of the dense models; it is not itself a mixture-of-experts variant. The model card lists four billion parameters and describes support for multilingual tasks. Performance statements in the model documentation are developer claims, and results for larger members of the family should not be assumed to apply to this model.
https://huggingface.co/Qwen/Qwen3-4B
Release and Availability
The Qwen team announced the Qwen3 family on 29 April 2025. Its release post lists the 4B dense model among the open-weight releases and links to distribution platforms including Hugging Face, ModelScope and Kaggle. The release is international; the sources do not describe a separate UK model.
https://qwenlm.github.io/blog/qwen3/
Thinking Mode and Budget Control
The model card documents enable_thinking as a switch in the tokenizer chat template. When enabled, reasoning is returned in a think block before the final answer; when disabled, that block is omitted. The documentation also explains that integrations expose this setting differently, so developers should follow the instructions for their chosen runtime.
https://huggingface.co/Qwen/Qwen3-4B
Deployment Options
The release post recommends SGLang and vLLM for deployment and lists local tools including Ollama, LMStudio, MLX, llama.cpp and KTransformers. These are documented options rather than a guarantee that every laptop can run every configuration. Before choosing a runtime, check its model support and the requirements for the configuration you intend to use.
https://qwenlm.github.io/blog/qwen3/