A broad model family
Qwen is the family of AI models developed by Alibaba Group’s Qwen Team, known in Chinese as Tongyi Qianwen. It covers not only large language models but models for images, audio, video, programming, and mathematics.
Rather than one AI called Qwen, it is a family of models differing in purpose and size. Names carry a suffix indicating use or generation — Qwen, Qwen-VL, Qwen-Audio, Qwen-Coder.
The main lines
- Qwen: language models for understanding and generating text
- Qwen-VL: models that understand text together with images and video
- Qwen-Audio, Qwen-ASR, Qwen-TTS: speech recognition and speech generation
- Qwen-Coder: models focused on writing, explaining, and fixing code
- Small models: lightweight versions that run on a laptop or a device
Generation names such as Qwen3 and Qwen3.5 also appear. Within one generation, models differ in supported input types, size, and how they are released, so read past the number to the suffix.
Hosted service versus open weights
The easiest way to try Qwen is through official web services such as Qwen Studio, with a conversational interface much like ChatGPT. Developers can build against the Alibaba Cloud API.
Some models are also published as weights on Hugging Face and can run on your own machine, which keeps data local but requires matching memory, GPU, runtime, and configuration. Not all Qwen models are open — some are available only as a hosted service.
Compared with DeepSeek and Llama
Qwen, DeepSeek, and Llama all appear as large model families that include open releases. Qwen is distinguished by the breadth of its purpose-specific lines across vision, audio, and code as well as text.
Choosing on benchmark position alone often produces a poor fit. Check quality in the language you need, the input formats required, the size that runs on your hardware, licensing, API pricing, and tool-use support. Names and terms change frequently, so read the current official documentation.