What it is
Qwen is a family of AI models made by Alibaba Cloud, the cloud computing arm of the Chinese technology company Alibaba Group. Like GPT, Claude, and Gemini, the Qwen models are language models: systems that have learned from huge amounts of text to predict and write words, and so can answer questions, draft text, write code, translate, and more.
What makes Qwen stand out is two things. The first is breadth: Alibaba has released hundreds of models under the Qwen name, from tiny ones that run on a phone to giants too large for anything but a data centre, and in versions for text, code, maths, images, video, and audio. The second is openness: many of these models are "open weight", which means anyone can download the finished model and run it on their own computer or server, free of charge, under a licence.
The name comes from Tongyi Qianwen, the Chinese name of the model Alibaba Cloud first showed on 11 April 2023 at its cloud summit in Beijing. At that launch Alibaba said it would build the model into its own products, starting with DingTalk, its workplace messaging app, and Tmall Genie, its smart speaker. On 3 August 2023 Alibaba Cloud took a step that set Qwen apart from most big companies at the time: it published Qwen-7B and Qwen-7B-Chat, with seven billion parameters each, for anyone to download from Hugging Face and from ModelScope, Alibaba's own model site. Companies with fewer than 100 million monthly users could use them commercially at no cost.
The generations, in brief
Model families change quickly, so treat this as a snapshot at the time of writing, with dates taken from Alibaba's own announcements where possible:
- Qwen (2023). The first open models, from 1.8 billion parameters up to 72 billion, plus early vision (Qwen-VL) versions.
- Qwen2 (June 2024) and Qwen2.5 (19 September 2024). Qwen2.5 came in seven sizes from 0.5 billion to 72 billion parameters, together with Qwen2.5-Coder for programming and Qwen2.5-Math for maths. The Qwen team called it possibly the largest open-source release in history.
- Qwen2.5-Max (late January 2025). A large model that Alibaba kept closed, released during the Lunar New Year holiday just after DeepSeek's breakthrough, with Alibaba claiming it beat DeepSeek-V3 on several tests.
- Qwen3 (29 April 2025). Six "dense" models from 0.6 billion to 32 billion parameters and two mixture-of-experts models, all under the Apache 2.0 licence. Alibaba said they were trained on about 36 trillion tokens and support 119 languages and dialects. Each can switch between a quick answer mode and a step-by-step reasoning mode, like the extended thinking found in other families.
- Qwen3-Coder (22 July 2025) and Qwen3-Max (September 2025). A large coding model, and Alibaba's biggest model at the time, with over a trillion parameters, available only through Alibaba's cloud.
- Qwen3.5 (February 2026) onwards. The open flagship, Qwen3.5-397B-A17B, handles text, images, and video. In August 2026 Alibaba released Qwen3.8, including a 27-billion-parameter model under Apache 2.0 and a downloadable version of its largest "Max" class model under a custom licence.
Reading a Qwen model name
Qwen names look like codes, but they carry useful information. "Qwen3-8B" means the Qwen3 generation with about 8 billion parameters. "Qwen3-30B-A3B" means 30 billion parameters in total but only about 3 billion active for each word, which is the mixture-of-experts trick explained on that page. "Coder", "Math", "VL" (vision-language), "Audio", and "Omni" (text, images, sound, and video together) tell you the speciality.
The word "Instruct" matters most. A base model has only done pretraining: it continues text, but it is not good at following requests. An Instruct model has had extra training, called fine-tuning, to follow instructions and hold a conversation. If you want a chat assistant, you want an Instruct (or chat) version. Base models are mainly for developers who plan to fine-tune them for their own purposes.
Open weights, closed models, and licences
"Open weight" is not the same as "open source" in the full sense. Alibaba publishes the finished model and a technical report, but not the complete training data, so you can use and adapt the model but you cannot rebuild it from scratch. The licences also vary, which is easy to miss:
- Many Qwen models, including the whole Qwen3 family and most of Qwen2.5, use Apache 2.0, a standard permissive licence that allows commercial use, changes, and redistribution.
- Some models used Alibaba's own Qwen licence. For example, the Qwen2.5 3B and 72B models were excluded from Apache 2.0. The 72B licence requires anyone whose product has more than 100 million monthly active users to ask Alibaba for permission, and asks anyone who uses the model to build or improve another AI model to display "Built with Qwen" or "Improved using Qwen".
- The large Qwen3.8 "Max" download from August 2026 has its own licence that, among other terms, requires companies earning over 50 million US dollars a year from selling AI models or AI assistants as a service to get a separate licence.
- The top models, such as Qwen3-Max and the "Plus" versions, are closed: you can only use them through Alibaba's cloud service or app.
Licence terms change between versions, so always read the licence file of the exact model you download.
Why it matters
It is one of the most widely used open model families in the world. A December 2025 issue brief from Stanford HAI and DigiChina reported that in September 2025 Qwen overtook Meta's Llama to become the most downloaded family of language models on Hugging Face, the main website where open models are shared. The same brief reported that since January 2025, new models built on Alibaba and DeepSeek models had been uploaded faster than those built on the major US and European families. Download counts are an imperfect measure, and the brief says so, but the direction is clear.
Other people build on it. Because the models are free to download and many come under a permissive licence, companies, universities, and hobbyists take Qwen models and fine-tune them for their own languages and tasks. DeepSeek itself published smaller "distilled" versions of its R1 reasoning model built on Qwen 2.5 and Llama 3. Researchers use Qwen too: the experiment described on the Temperature page, in which one prompt gave 80 different answers at temperature zero, was run on Qwen3-235B. The big cloud providers offer it as well: Amazon added four Qwen3 models to its Bedrock service in September 2025.
The size range suits real budgets. A small Qwen model can run on an ordinary laptop, with no internet connection and no fees, while the largest compete with the best closed models on many public benchmarks. That lets a business choose the smallest model that does the job well enough, which is often much cheaper and faster than always calling the biggest one.
It raises real questions about trust and politics. Qwen is made by a Chinese company, and models that are legally offered in China must follow Chinese content rules. That affects how the models answer some political questions, and it has become part of the argument between the United States and China over AI. Anyone choosing a model should understand this, and the examples below show it playing out.
Real-world examples
Airbnb's customer service agent (October 2025 to April 2026)
In an interview with Bloomberg in October 2025, Airbnb's chief executive Brian Chesky was asked why Airbnb had not yet connected its app to ChatGPT. As reported by Fortune on 21 October 2025, he said Airbnb was "relying a lot" on Alibaba's Qwen model, calling it "very good" and "also fast and cheap", and said Airbnb used OpenAI's latest models but typically not much in production, because faster and cheaper models were available. Fortune reported that Airbnb's AI customer support agent, which had been rolled out to all US English-language users in May 2025, drew on 13 different models from OpenAI, Alibaba, Google, and open-source providers.
The comments were widely read as a sign that open Chinese models had become a practical choice for large Western companies. They also drew political attention. On 29 April 2026, the chairmen of the House Select Committee on China and the House Committee on Homeland Security announced a joint inquiry, writing to Airbnb for more information about its reliance on Qwen for customer service, citing national security and data security concerns. The same inquiry covered Anysphere, the maker of the coding tool Cursor, over a model reportedly built on one from Moonshot AI, the Chinese maker of Kimi. Nextgov reported that Airbnb did not immediately respond to requests for comment. The story shows both why businesses pick Qwen (cost and speed) and the questions they may then have to answer.
Singapore's national language model switches to Qwen (November 2025)
AI Singapore, the country's national AI programme, has built a model family called SEA-LION since 2023 to handle Southeast Asian languages that big Western models often handle poorly. Earlier versions were built on other open families, including Llama. On 24 November 2025 Alibaba Cloud announced Qwen-SEA-LION-v4, built on Qwen3-32B and trained further on more than 100 billion tokens of Southeast Asian language text, covering languages including Burmese, Filipino, Indonesian, Malay, Tamil, Thai, and Vietnamese. According to the announcement, it ranked first among open models under 200 billion parameters on the SEA-HELM leaderboard, which tests these languages, and a compressed version can run on a laptop with 32 GB of memory. The model is free to download.
This example shows the most common way Qwen is used: not as a finished chatbot, but as a foundation that someone else adapts. A national programme could take an existing open model, add its own language data, and publish the result, without training a model from nothing. That ranking claim came from Alibaba and AI Singapore, so check the current leaderboard yourself if it matters to your choice.
The Qwen app (November 2025)
Most of Qwen's reach outside China comes through downloads and cloud services, but inside China Alibaba also offers it directly to consumers. On 17 November 2025 Alibaba launched the Qwen app in public beta as a free personal assistant for users in China, powered by Qwen3, with deep research, coding help, camera features, voice calls, and the ability to turn research into a slide presentation. Alibaba said on 25 November 2025 that it had passed 10 million downloads in its first week. Alibaba said an international version would follow later; check its availability in your country at the time you read this.
An everyday case: the translator's private glossary
Here is an illustrative scenario, not a real company. Mei runs a small translation business that handles technical manuals for local manufacturers. Her clients will not let their documents go to an outside AI service. She installs a free tool on her office computer and downloads a mid-sized Qwen Instruct model. It runs entirely on her machine, so nothing leaves the building. It isn't as polished as the best cloud models, but for first drafts of routine translation it is good enough, and she checks every page herself. For a novel or a legal contract she still does the work by hand. She also notes that the model's answers on anything political are not something to rely on, which does not matter for pump manuals but would for other work.
Try it yourself
This exercise shows you how model size affects quality, using two free Qwen models on your own computer. You don't need to be technical, and you don't need an account. It uses Ollama, a free program for running open models locally; the Parameters page explains more about running models on your own machine.
Hardware at the time of writing: on Ollama's Qwen3 page, the 1.7-billion-parameter model is a 1.4 GB download and the 8-billion model is 5.2 GB. The small one runs on most laptops from the last few years. For the 8-billion model, a computer with 16 GB of memory is a comfortable starting point; with less it may be slow or fail to load. You'll also need free disk space for the downloads. Sizes and requirements change, so check the Ollama page first.
- Install Ollama. Go to ollama.com, download the version for Windows, Mac, or Linux, and install it like any other program.
- Open a terminal. On Windows, open the Start menu and type "PowerShell". On a Mac, open the "Terminal" app. It's a window where you type commands.
- Download and start the small model. Type this and press Enter. The first time, it downloads the model, which may take a few minutes: > ollama run qwen3:1.7b
- Give it a task. When you see a prompt, paste this: > A bakery sells muffins for $3 each or 4 for $10. I need 11 muffins. What is the cheapest way to buy them, and how much will I pay? Then write a two-sentence friendly sign for the shop window advertising the 4-for-$10 deal. The cheapest answer is two sets of four plus three single muffins, for $29. Note whether the model gets the maths right and how natural the sign sounds. Qwen3 may show its reasoning first; that's its thinking mode.
- Close it by typing /bye and pressing Enter.
- Now try the bigger model with exactly the same prompt: > ollama run qwen3:8b
- Compare. Which one got the sum right? Which sign would you actually put in a window? How long did each take? Try one more test in both, in a language you know other than English, or on a topic from your own work.
- Ask each model about itself. Type "Which model are you, and what is your knowledge cut-off date?" Smaller models often get details about themselves wrong, which is a good reminder that models can hallucinate about anything.
- Write down your rule. Finish this sentence: "The small model is good enough for ___, but I'd use the bigger one for ___." That's the same trade-off companies make when they choose a model size.
If your computer is too old for this, you can instead try the same prompt on two different models on Qwen's own chat website, chat.qwen.ai, if it offers a choice of models when you visit. At the time of writing it may ask you to sign in, and anything you type there is processed on Alibaba's servers, so don't paste anything private.
Common mistakes and misunderstandings
Thinking "Qwen" is one model. It is a family of hundreds of models of very different sizes, skills, and licences. "I tried Qwen and it was poor" might mean someone tried a 0.6-billion-parameter model built for phones. Always note the exact name and size.
Assuming every Qwen model is free for any use. Many are under Apache 2.0, but some older and larger ones had Alibaba's own licence with conditions, the largest download from 2026 has a custom licence with revenue limits, and the top models are closed. Read the licence of the exact model you use, and ask a lawyer if the stakes are high.
Mixing up running it yourself with using Alibaba's service. When you download a Qwen model and run it on your own computer or a cloud provider you trust, your data stays there. When you use the Qwen app, chat.qwen.ai, or Alibaba Cloud's service, your data goes to Alibaba. The Stanford HAI and DigiChina brief made the same point: running the model locally or through a trusted provider greatly reduces the data risk. The Privacy page covers what to check.
Forgetting the political filter. Research published in March 2026 by Casademunt and colleagues found that Qwen3 models often give false answers about subjects such as Falun Gong and the 1989 Tiananmen protests, while sometimes answering correctly, which suggests they know more than they say. This comes from the rules Chinese developers must follow, and is not unique to Qwen; the DeepSeek page covers the wider picture. Fine-tuned versions made by others may behave differently. For history, politics, or current affairs involving China, check other sources. Western models have their own biases too, so no model should be your only source on contested topics.
Picking a base model for chat. If a downloaded model rambles or continues your question instead of answering, you may have a base model. Look for "Instruct" or "chat" in the name.
Treating benchmark claims as independent facts. Alibaba, like every model maker, publishes scores from tests it chose. Others do the same for their own models. Check independent results, and above all test on your own tasks, as the Evals page explains.
Expecting it to stay the same. New Qwen versions have arrived every few months since 2023, with names and licences shifting each time. Anything in this article about versions, sizes, or terms may be out of date by the time you read it, so check the Qwen team's pages.
Related elements
- DeepSeek: China's other best-known open model maker, with the wider discussion of Chinese content rules.
- Llama: Meta's open-weight family, which Qwen overtook in Hugging Face downloads in 2025.
- Mistral: the leading European maker of open-weight models.
- Kimi: Moonshot AI's models, another Chinese open-weight family.
- GLM: Zhipu's family, part of the same wave of Chinese open models.
- Parameters: what the "B" in names like Qwen3-8B means, and how to run models on your own computer.
- Mixture of Experts: the design behind names like Qwen3-30B-A3B.
- Fine-tuning: how others turn a Qwen base model into their own specialist model.
- Temperature: includes the experiment on Qwen3-235B that produced 80 different answers at temperature zero.
- Benchmarks: how to read the scores every model maker publishes.
- Privacy: what to check before sending data to any AI service.