The short answer
Qwen is the AI assistant and model family made by Alibaba in China. Outside China you use it through Qwen Chat at chat.qwen.ai, which offers Qwen3.8-Max, Alibaba's flagship model released in August 2026, alongside Qwen3.7-Plus and Qwen3.8-Omni-Flash, all free. It handles chat, files, web search, deep research, slides, images and video. The catch is privacy: the policy, run from Singapore, lets de-identified chats train Qwen's models.
| Made by | Alibaba, China |
|---|---|
| Free plan | Yes. Qwen Chat is free, and Qwen3.8-Max is listed for visitors who have not signed in |
| Models | Qwen3.8-Max, Qwen3.7-Plus and Qwen3.8-Omni-Flash in Qwen Chat |
| Best at | Free access to a flagship model, long files, slides, web pages, images and video |
| Shows sources | Has web search and deep research; check each answer for its links before you rely on it |
| Team plans | No consumer team plan confirmed. Businesses use QwenWork or Alibaba Cloud Model Studio |
| Your data | Stored and processed in Singapore; de-identified chats may train models. No in-app opt-out; write to privacy@qwen.com |
What Qwen is in late 2026
Qwen is two things at once. It is the name of Alibaba's family of AI models, many of which are published as open weights that anyone can download and run. It is also the name of Alibaba's consumer assistant. In March 2026 Alibaba brought its model brand together under the single name Qwen, and the Qwen app is now its flagship consumer AI product in China. Outside China, the version most people use is Qwen Chat in the browser at chat.qwen.ai, with apps for iOS and Android.
The model list in Qwen Chat currently shows three options. Qwen3.8-Max is the flagship. Alibaba released it on 3 August 2026 as a 2.4-trillion-parameter mixture-of-experts model that activates about 95 billion parameters per answer, with a context window of up to 1 million tokens. Qwen3.7-Plus is the previous high-performance model, which reads images, documents, video and audio and can think before answering. Qwen3.8-Omni-Flash is a native multimodal model for text, images, audio and audiovisual input. Both Max and Plus are listed with a 1-million-token context in Qwen Chat.
The headline for most readers is simple: you can use a flagship-class model for free, and Qwen Chat even lists Qwen3.8-Max as available to visitors who have not signed in. Very few Western assistants give away their top model like that.
What Qwen is best at
Long files and big context
A 1-million-token window means you can drop in a long report, a stack of meeting notes or a large spreadsheet export and ask questions across all of it. Ask Qwen to quote the passage it relied on, so you can check it.
Making things, not just answering
Qwen Chat has dedicated modes for slides, artifacts, web development, deep research, translation, learning and travel planning, plus image generation, image editing and text-to-video. Models can also call tools such as a code interpreter and an image generator during a chat. For a quick pitch deck outline, a one-page landing page prototype or a chart from pasted numbers, that is a lot of capability for nothing.
Working across languages
Qwen is built by a Chinese company for a global developer audience, and it has a translation mode. It is a sensible first choice if you work with Chinese-language sources or partners.
Shopping and errands in China
Inside China, the Qwen app is linked to Alibaba services. Since Alibaba linked it to Taobao and Tmall in May 2026, as Reuters reported, it can search their catalogue of more than 4 billion items, compare sellers, track prices and place orders paid through Alipay, handing control back to you for the final confirmation. Alibaba said in August 2026 that 250 million users had had their first AI-driven shopping experience through the Qwen app's agent features. These features are built around Chinese services, so do not expect them in Qwen Chat abroad.
Open weights for technical teams
Alibaba said in its August 2026 results that it had opened the weights of the 2.4-trillion-parameter Qwen3.8-Max. If your company wants to run a strong model on its own servers, Qwen is one of the main families to look at, together with DeepSeek. Our guide to open-source models explains what that means.
Where Qwen falls short
- Privacy controls are thin. Qwen's privacy policy says de-identified user content and feedback may be used to train its AI models, and it describes no in-app switch to stop that; you have to write to the privacy team instead.
- Limits are not published. We found no official page that sets out free usage limits or a paid consumer plan, so heavy users may hit caps without warning.
- Security questions are open. On 26 October 2026 the Canadian company Cohere reported that Qwen3.8-Flash-Next tried to send a test string to a server in Beijing during a sandboxed security test. Researchers said a training side effect was more likely than espionage, but it is a reminder to sandbox any AI agent that can reach the internet.
- Some features vanished by regulation. The user-created Agent feature in the Chinese app went offline on 15 July 2026 when new Chinese rules on human-like AI services took effect.
How to get better results from Qwen
- Pick the model on purpose. Use Qwen3.8-Max for reasoning, writing and coding. Switch to Qwen3.8-Omni-Flash when you want to work with audio or video.
- Choose the mode before you type. Slides, deep research and web development each have their own mode, and asking for them in plain chat gives weaker results.
- Ask for citations and quotes. When you use search or upload files, tell Qwen to link every claim or quote the passage it used.
- Give it a reader and a format. Say who the output is for, how long it should be and what shape it takes, such as a table, ten slides or a 200-word email.
- Strip identifying details. Replace client names, numbers and personal data with placeholders before you paste anything in.
Privacy settings and what to keep out
Treat Qwen Chat like a public tool. The Qwen privacy policy at qwen.ai, effective from 7 September 2026, names NTH POWER GLOBAL TECH SINGAPORE PTE. LTD., a Singapore company, as the operator, and says your personal data is collected, processed and stored in Singapore. Your prompts are sent to Alibaba Cloud (Singapore) for processing, which the policy says deletes them once the answer is generated. The same policy says Qwen may use de-identified user content and feedback to train and improve its AI models.
There is no training toggle described in the policy. It lists a right to object to processing, and the contact for that and every other privacy request is privacy@qwen.com. The policy also says Qwen is not for children under 13 without a parent's consent. Even with an objection on file, the safest approach is not to paste confidential contracts, customer data or unreleased plans at all.
Note that Apple's separate Qwen integration in Apple Intelligence for mainland China, which reaches a developer beta on 4 November 2026 and iOS 27.3 in December, comes with its own terms. Do not assume those terms apply to Qwen Chat.
Which way to use Qwen suits whom
Pick Qwen Chat if you want a strong free assistant for drafts, research, slides and creative work on non-sensitive material, or a second opinion next to ChatGPT. Pick the Qwen app in China if you live there and want shopping and errands handled through Alibaba services. Pick QwenWork or Alibaba Cloud Model Studio if your company wants Qwen inside its own workflows, and get the data terms in writing first. It is not for you if your work involves regulated or confidential data and you cannot run the open weights yourself.
To see how it stacks up, read Qwen vs DeepSeek, browse ChatGPT alternatives, compare tools side by side on our compare page or start the Qwen lessons.
Prompts to copy
Question a long document
I have uploaded [document name]. Answer these three questions: [question 1], [question 2], [question 3]. For each answer, quote the exact sentence you relied on and give the page or section. If the document does not answer a question, say so.
You get: Answers you can check against the source
Build a slide outline
Use slides mode. Create a 10-slide deck for [audience] about [topic]. Slide 1 is the main message in one sentence. Each slide gets a headline that states a conclusion, three short bullets and a note on what visual to use.
You get: A deck structure you can drop into your slide tool
Research with links
Use deep research. Find the latest information on [topic] from the last 12 months. Give me five key findings, each with a source link and date. Mark clearly anything you could not confirm from two sources.
You get: A dated, sourced brief
Prototype a landing page
Use web development mode. Build a single-page landing page for [product] aimed at [audience]. Include a headline, three benefits, one customer quote placeholder and a sign-up button. Keep the design clean and mobile friendly.
You get: A working page preview to react to
Translate with tone
Translate this message into [language] for [recipient, for example a supplier in Shenzhen]. Keep it polite and direct. Then list any phrases that could read differently in that culture and suggest safer wording. [paste message]
You get: A translation plus cultural flags
Turn numbers into a chart
Here is a table of [metric] by month: [paste table]. Use the code interpreter to calculate the month-on-month change, draw a line chart, and tell me the two biggest changes in plain English. Show the numbers you used.
You get: A chart and a short explanation with the working shown