An independent guide to Qwen
Qwen AI:
what it is and how to use it.
Qwen is Alibaba’s family of AI models and the name used for its AI assistant. The assistant gives you a place to chat, work with files and use creative tools. Developers can access supported models through hosted APIs or run selected downloadable releases themselves.
ChinaAI is an independent publication, not Alibaba or Qwen. These links open the official services.
By China AI Chat Team · Guide updated September 27, 2026 · Tests and source-check dates below · Our research method
Qwen Chat, Model Studio or a downloadable model?
For everyday use, start with the official assistant. Its web destination is chat.qwen.ai, currently titled Qwen Studio. Alibaba Cloud Model Studio is a separate developer platform. The similar names do not mean they share the same interface, account benefits or billing.
USE THE ASSISTANT
Chat and create
Ask questions, supply source material and use the tools shown in your account. You do not need to build an application or manage a model server to start.
BUILD AN APPLICATION
Use a hosted API
Call a model from your own software. Choose the provider, model and deployment region, then configure its credentials and usage budget.
RUN A MODEL YOURSELF
Explore model releases
Choose a published checkpoint that fits your hardware and task. Check its license, required software and memory before downloading.
For a concrete Windows walkthrough, see how to run Qwen locally with Ollama.
The official Qwen organization identifies Alibaba Cloud as the developer and links to Qwen Chat. Alibaba documents the separate developer setup in its Model Studio quickstart.
What can Qwen AI do?
Qwen’s official product overview describes writing, reasoning, search, document analysis, media understanding and creative tools. These are product capabilities, not a guarantee that every Qwen model or third-party app supports every task.
Choose the input and output first. Asking a question about a picture is an image-understanding task; changing that picture requires an image-editing tool. A model that accepts images does not necessarily generate them.
For a short explanation, draft or summary, start with the official Qwen assistant and check its answer against your source. For a more specific task, choose a guide below.
| Your task | Start here | What the guide helps you check |
|---|---|---|
| Research a question across sources | Use Qwen Deep Research and check the report | Follow a real Advanced-mode task, inspect important factual errors and preserve the exported sources. |
| Create a short video | Generate and download a Qwen video | See the official controls, original approximately five-second clip and where its camera and lighting instructions were missed. |
| Create a presentation | Create, check and download Qwen Slides | Inspect the first deck and an assisted correction, check remaining chart errors, and examine both PDF exports. |
| Read a PDF | Upload a PDF and verify the answer | Check extracted facts, page references and calculations against the original document. |
| Create a new image | Generate images from a prompt | Inspect composition, lettering and output dimensions in two original examples. |
| Change an existing picture | Edit an image in Qwen Chat | Compare the requested change, details that should remain and the exported resolution. |
| Work on project files | Set up Qwen Code | Connect the terminal coding agent to QwenCloud or another provider, then inspect a tested feature addition, an earlier repair and the limits of each test. |
| Run a model on your computer | Run Qwen locally on Windows with Ollama | Follow the tested setup and inspect its CPU results, memory observations and incorrect answers. |
| Choose a model | Understand Qwen models and releases | Distinguish hosted identifiers, downloadable weights, supported inputs and licenses. |
| Connect your application | Use a Qwen API | Choose the provider and inspect our three tested QwenCloud requests, with separate Model Studio setup instructions. |
| Understand the cost | Check Qwen pricing and free-access rules | Separate consumer access from API charges and estimate a defined workload. |
How to get started with Qwen
- Open the official chat or app. Follow the account steps it presents. For a download, start at Qwen’s official download page to reach the appropriate product for your device.
- Choose one small task. State what you want back: a short answer, a table, a code change or an edited image. Check that the chosen tool accepts your input.
- Supply the material and constraints. Attach the intended file, name the audience or purpose, and identify what must remain unchanged. Check the attachment before sending.
- Inspect the answer or file. Verify facts against sources, run generated code, or compare image details and dimensions. Revise a specific missed requirement rather than judging only by presentation.
For a first task, try this prompt with a short set of notes. It is a suggested template, not an additional test:
Read the notes below and create a table with three columns: task, owner and deadline. Use only information present in the notes. Write “not specified” for missing owners or dates. Do not add tasks. After the table, list any conflicting deadlines. Notes: [paste non-sensitive notes here]
To create a new image from text, follow our Qwen image-generator guide. It includes two first-result tests, exact prompts and a check of the output dimensions.
For an image edit, describe both the intended change and the features to preserve. The before-and-after example below shows why a visual check still matters.
Is Qwen AI free?
Free consumer access is available. Qwen’s official product page describes the assistant as free to use. However, the official Qwen Android listing also shows in-app purchases. Check the app, region and account offer you are actually using; free access is not a promise of unlimited use or every feature at no charge.
Our image-editing run completed with the access already available to us, without a new purchase or API top-up. We did not establish a universal free-editing quota.
Hosted APIs have their own billing. Alibaba Cloud publishes Model Studio rates, and eligibility for trial quotas depends on the applicable offer. Another company hosting a Qwen model can charge different rates. Running a downloadable model yourself also requires computing resources, even when the weights are available without a download fee.
Use our Qwen pricing guide for separate QwenCloud and Model Studio rates, Token Plan subscriptions, free-quota conditions and worked examples. Choose the product and workload before estimating cost; “Qwen” alone is not a pricing plan.
What our Qwen tests found
Each guide links its original inputs, results, date and settings. The tests use different setups: consumer chat, direct QwenCloud requests, an API-backed coding agent and a local model.
PDFs: check the page references
In three first-answer tests, the selectable-text brief and order table passed their checks. The clean image-only brief returned the requested facts but gave two wrong page references. A plausible answer still needs checking against the file.
Image generation: inspect the dimensions
Both first outputs contained the requested objects and poster text. The poster measured 1728 × 2368, slightly different from the selected 3:4 ratio. Two images cannot show how reliably other prompts will work.
Qwen Code: inspect the change, then verify it
On September 25, Qwen Code with QwenCloud’s qwen3.8-max added a booking-summary feature in our restricted test setup. All 19 predefined check groups passed when we ran them independently, and the agent session completed successfully. The agent was not given permission to run the tests.
Our September 21 test with Together’s Qwen3.5-9B made a one-line repair that passed 13 independent checks, but its session ended with a tool-call error. These different tasks and providers do not establish which model is better.
Inspect the feature test and downloadable project → · Read the earlier repair and limitation →
Local Qwen: a working setup can give wrong answers
We made nine CPU requests with Qwen3.5:9b on Windows through Ollama, with thinking disabled. Extraction passed all three runs. Each summary omitted a required detail, and every arithmetic answer was wrong. Those results apply to that configuration.
QwenCloud API: three small integration checks
On September 25, Qwen3.8-Max returned the expected date in ordinary and streamed requests and the exact fields in a strict JSON extraction. The three requests consumed 146 tokens, matching the free-quota dashboard. These are connection and output-format checks, not a model-quality benchmark.
Inspect the QwenCloud requests and evidence → · Understand the recorded usage and cost calculation →
Deep Research: a useful conclusion can contain wrong details
One Advanced-mode report on September 25 correctly withheld approval for an unspecified USB-C docking setup, but misstated generation names and charging requirements. The guide includes eight predefined checks, a source audit and the original PDF.
Inspect the research workflow, source audit and original report →
Slides: readable design still needs source checks
On September 25, one museum-pilot brief produced a downloadable six-page, image-based PDF. The first deck retained the headline contribution and conditional checks, but changed two attendance values and drew misleading charts. We did not receive a native editable PowerPoint file in that workflow.
In one assisted follow-up on September 26, we supplied the correct values. Qwen restored the attendance labels and financial breakdown, but its attendance chart still misrepresented the figures. The revised export remained an image-based PDF.
Inspect the first slides and original PDF → · See the correction trial and remaining errors →
Video: check the camera movement and lighting
In one first-output video test, the suitcase and handle remained recognizable in our frame review, but the clip missed the requested camera sequence and soft studio lighting.
The guides preserve failures alongside successful outputs. See how we research and test for our evidence standards.
An image edit: what changed beyond the color
In one Qwen Chat edit on September 18, 2026, we asked for a fictional blue mug to become terracotta orange while keeping its shape, framing and surroundings unchanged. We kept the first result without a retry.
The color changed, but some texture, edges and highlights changed too. The exported image also had half the original width and height. Keep the original and compare details at full size before using an edit in a design.
This is one example, not a measure of general editing accuracy. The full image-editing guide contains the exact prompt, recorded tool labels, original files and troubleshooting steps.
Compare Qwen with Kimi, DeepSeek and ChatGPT
Start with the task you need to complete, then inspect the answers and checks. These are dated tests of official consumer web apps, including dedicated slide workflows in the Kimi comparison; they are not API benchmarks.
Qwen vs Kimi
Three first-run app tasks: a sourced report brief, a fictional revised quote and a six-slide board packet. Kimi followed the report more closely; both reconciled the quote; neither first deck was presentation-ready. The page shows original answers, slide captures and test limits.
Qwen vs ChatGPT
Four short English tasks, two requests per task in each app. All completed answers passed our checks. One Qwen request produced no answer, and another returned a passing answer alongside a blank alternative. These session outcomes do not establish general reliability.
Qwen vs DeepSeek
Four short text tasks, three requests per task in each app. DeepSeek retained all required summary details in three runs and Qwen in one. Both passed the other task checks in every run. This small collection does not establish an overall winner.
These comparisons use different prompts, settings and workflows. Their results should not be combined into a multi-model ranking. For help choosing a study, use our Chinese AI comparison guide organized by task.
How to understand Qwen model names
The Qwen family includes releases for different tasks and inputs. A general language model, a vision-language model that answers questions about images, and an image-generation model serve different purposes. An assistant can connect several tools around them.
For example, the official catalog includes Qwen3.8 releases and Qwen-Image releases. These names identify model lines, not identical feature sets or account allowances. Use the exact model card or hosted model reference to check supported inputs and outputs.
- For the assistant: check the selected mode, available tools and limits in the current interface.
- For an API: record the provider, model identifier and region. Alibaba’s API-key documentation explains the region-specific setup.
- For a local model: inspect the exact checkpoint, license, memory requirements and supported runtime. A downloadable release does not mean every laptop can run it well.
Our Qwen model guide covers the model choices in more detail. The API guide covers developer access. Model Studio also hosts non-Qwen models, so check the developer as well as the platform name.
Common questions about Qwen AI
Is Qwen owned by Alibaba?
Yes. The official Qwen organization describes the model family as developed by Alibaba Cloud, and its GitHub organization identifies the Qwen team at Alibaba. Qwen is the product and model brand; ChinaAI is an independent guide.
Can I use Qwen in English?
Yes. Its official product site is available in English, and our text comparisons, PDF tasks and image prompts used English. Those examples do not establish accuracy across all subjects. Specify English for the output when it matters to your task.
Is Qwen open source?
Qwen publishes downloadable models and code, but you should check the license of the exact release you intend to use. A hosted model name is not proof that its weights are downloadable. Consumer-assistant access, API access and permission to run a model are separate questions.
Is Qwen safe to use?
Consider the service and the material you plan to share. Check the privacy policy and account controls in your chosen app before uploading confidential files; check what an API provider retains before integrating it. We have not audited those services. A downloadable model does not make a hosted chat session private, and a generated answer or image still needs review.
Is Qwen better than ChatGPT or DeepSeek?
The task-specific comparisons above show the answers and session limits. They do not support an overall winner.
Can I use Qwen chat on ChinaAI?
Live visitor chat is not active here. The ChinaAI chat directory provides official links, a local question draft you can copy and clearly labelled recorded examples. Writing a draft does not send it to a model. Open the official Qwen service for a new live answer.
Sources, test dates and corrections
The page was reorganized on September 27, 2026 to bring getting-started steps and access costs ahead of the test summaries. No new product tests were performed for this edit. The Slides follow-up summary was added September 26. Product descriptions were reviewed September 21, with the official product overview and model catalog spot-checked September 22. The original linked experiments were performed September 18–25; the Slides correction trial took place September 26. Each guide records its exact conditions and limits. Provider descriptions and suggested tasks are distinguished from our observations. How we research · Report a correction.