Qwen

Qwen is a family of open-weight large language models from Alibaba Cloud (the name is short for Tongyi Qianwen, 通义千问). Like a GPT , each Qwen model is a decoder-only transformer trained to predict the next token. The part that matters in practice is the open-weight part: Alibaba publishes the actual trained parameters , much of the family under the permissive Apache 2.0 license, so anyone can download a Qwen model and run it on their own hardware instead of renting it through someone else’s API.

The family comes in generations (Qwen, Qwen1.5, Qwen2, Qwen2.5, Qwen3) and a wide spread of sizes, from tiny half-billion-parameter models that fit on a phone up to very large mixture-of-experts models in the hundreds of billions. There are specialized branches too: Qwen-Coder for programming, Qwen-VL for images-plus-text, Qwen-Audio for sound, and reasoning-focused variants. Because the weights are open and even the small sizes are genuinely capable, Qwen has become one of the most common starting points for running models locally: it ships in GGUF form for llama.cpp and Ollama, and it’s a frequent base for LoRA fine-tunes that teach it a narrower job or a particular voice.

If GPT is the closed model you reach through an API, Qwen is the open one you can actually hold: download it, inspect it, fine-tune it, run it offline. That difference is most of why it turns up so often in local-inference and hobbyist projects, where the few flagship Qwen models that stay API-only matter far less than the dozens you can just pull down and run.