Qwen3 is Alibaba’s flagship family of open large language models — a full series of dense and Mixture-of-Experts models built for reasoning, coding, agents, and multilingual work.
Qwen3 is developed by the Qwen team at Alibaba Cloud and hosted on GitHub with around 28,000 stars; the repository is the official home of the model series, with releases continuing into 2026. The family spans compact dense models from 0.6B up to 32B parameters and MoE flagships such as a 235B-total-parameter model that activates only a fraction per token. Its signature design is hybrid thinking: a single model can switch between a deliberate thinking mode for hard math, code, and logic, and a fast non-thinking mode for everyday chat. Qwen3 models support 119 languages and dialects and are released as open weights.
The flagship Qwen3-235B-A22B activates roughly a tenth of its parameters per token, giving flagship-class quality at a fraction of the serving cost, while the dense Qwen3-32B targets single-GPU deployments. Follow-up releases in the series added dedicated variants for coding and other specialised work. Native context windows reach 128K tokens, and the models are integrated with mainstream inference engines such as vLLM and SGLang as well as local runtimes.
Qwen3 is free: the model weights are publicly released for download and self-hosting, and the GitHub repository provides the official documentation and deployment guidance. Alibaba Cloud also offers Qwen models through its paid cloud API if you prefer a hosted endpoint.
Developers and researchers who want frontier-class open-weight models — for local deployment, fine-tuning, agent building, or as the reasoning core of a product — without depending on a closed API.