qwenlm/qwen3

Qwen3 is a family of large language models, the kind of AI that powers chatbots and code assistants, developed by the Qwen team at Alibaba Cloud. The repository is the public home for the model family: it points to downloadable model checkpoints on Hugging Face and ModelScope, hosts a demo and chat site, and links to documentation that walks through how to use the models. The README describes both dense models and Mixture-of-Experts models in a range of sizes from 0.6B up to 235B parameters, with an updated "Qwen3-2507" generation that comes in Instruct and Thinking variants. Instruct is tuned for general chat, while Thinking is tuned to spend extra effort on harder reasoning tasks like math, science, and coding. A notable feature is the ability to switch between thinking mode and a faster non-thinking mode, plus support for very long inputs, 256K tokens by default and up to 1 million tokens. The documentation outlines several common ways people actually use these models: running them locally on CPU or GPU through tools like llama.cpp, Ollama, and LM Studio; deploying them at scale with SGLang, vLLM, or TGI; shrinking them with quantization techniques like GPTQ and AWQ to make GGUF files; and fine-tuning them with Axolotl or LLaMA-Factory. The README also notes Qwen3 supports more than 100 languages. You would use this when you want an open-weight LLM you can run yourself, for a chatbot, a coding helper, an agent that calls external tools, or any application where you do not want to depend on a closed API. The supporting code in the repo is primarily Python, and the full README is longer than what was provided.