What is Qwen? Alibaba's free AI models, explained

You've probably seen 'Qwen' pop up in AI news and wondered what it actually is. The name appears alongside ChatGPT and Claude in capability comparisons, it's downloaded by millions of developers, and it comes from a Chinese tech giant you might not have thought was a major AI player. Here's the plain-English version — what Qwen is, why it matters, and the catch you should know about.

Most famous AI tools — ChatGPT, Claude, Gemini — are things you use online through a company's website or app. The company controls the model and charges you (directly or via ads) to use it. Qwen is mostly different: it is open-weight, meaning Alibaba publishes the actual model files for you to download. Once you have them, you can run the AI on your own computer or server, without asking Alibaba's permission or paying per use. You can also modify the model — retrain it on your own data, tweak its behaviour, or build it into a product you sell.

The Apache 2.0 licence is the legal bit that makes all of that allowed. It is one of the most permissive software licences in existence — essentially 'do what you want with this, including commercially, as long as you keep the licence notice'. That is why businesses love it: no per-token bill, no revenue-sharing clause, no vendor lock-in. Developers and companies can build on Qwen the same way they build on open-source databases or operating systems — it just works, and nobody can take it away mid-project. This freedom is the main reason Qwen has been downloaded approaching a billion times and reportedly accounts for over half of all open-source AI downloads.

Alibaba released the Qwen3.6 family in April 2026, and the headline result was surprising: the Qwen3.6-27B model — 27 billion parameters, free to download — reportedly outperforms Alibaba's own previous flagship, which was nearly 400 billion parameters. That's the AI equivalent of a small, efficient car beating a massive truck on the road course. The reason is better architecture: Qwen3.6 uses more efficient engineering ('Mixture-of-Experts' on some versions) and better training, so it needs less raw scale to hit the same results. What this means for you: smaller models that can run on a decent laptop or a cheap cloud server are now competitive with much larger ones that needed a rack of expensive GPUs. The gap between 'what a big company can afford' and 'what a developer can run at home' is closing fast.

Here is a quick overview of the main Qwen3.6 models:

The open models cover most people's needs. The closed one is aimed at businesses that need the absolute best and are happy to pay per API call.

There is a twist. In April 2026, Alibaba released its most powerful Qwen model yet — Qwen3.6-Max-Preview — but kept it paid and online-only. You cannot download it; you have to pay to use it through Alibaba Cloud's service. This is the first time a flagship Qwen model has been closed. It follows a pattern you may recognise: give away the good stuff to build a following, then charge for the very best. It is the same approach that made Spotify free for most users and paid for the extra features.

There is also one other thing to be aware of: because Qwen is made by a Chinese company, some governments and businesses in sensitive industries are cautious about using it, particularly for data that cannot leave certain jurisdictions. For most people experimenting with AI, building a side project, or running a business that does not handle sensitive government or financial data, that caution does not apply — and a free, capable, genuinely open AI model is exactly as useful as it sounds.