Qwen Deep Research is Qwen’s multi-step research workflow for turning a broad question into a structured, source-linked report. Instead of producing one immediate response, it can clarify the request, plan the investigation, search the web repeatedly, combine evidence, and generate a longer synthesis. You can begin from the official Qwen research page, but availability, mode names, usage allowances, and export controls can differ by account, platform, region, and rollout stage.
Last verified: October 7, 2026. The public Qwen pages, official announcements, policies, Model Studio documentation, and Alibaba-NLP open-source materials were checked for this guide. An authenticated consumer research job was not submitted, so account-specific interface controls are not presented as firsthand test results.
What Qwen Deep Research is—and which product you mean
The name can refer to three related but separate routes. They should not be treated as one product with the same interface, price, region, model, or privacy terms.
| Route | What it is | Who it is for | How you access it |
|---|---|---|---|
| Qwen Deep Research in Qwen | A hosted consumer research experience that plans searches and produces a cited report. | Students, analysts, marketers, researchers, writers, and general users. | The official Qwen website or another currently supported Qwen interface. |
qwen-deep-research in Model Studio | A managed developer model with a native research workflow and streaming responses. | Developers building research features into applications or internal tools. | Alibaba Cloud Model Studio with a supported regional API key and native interface. |
| Tongyi DeepResearch | An Apache-2.0 open-source research model and agent framework. | Teams that want to inspect, customize, or operate a research stack themselves. | The official Alibaba-NLP repository and model weights, plus the required search and processing tools. |
Similarity in naming does not prove that the consumer Qwen feature uses the exact Model Studio model or the exact open-source Tongyi architecture. No such relationship should be assumed without a current official statement.
Qwen’s November 2025 announcement used the name Qwen DeepResearch 2511 for an update that introduced Normal and Advanced modes, document and image input, more control over report structure, and improved citation handling. “2511” should be treated as a historical release label unless it still appears in the current interface—not as proof that it remains the newest version.
Qwen Search vs Thinking vs Deep Research
Deep Research is most useful when the answer depends on several sources, comparisons, changing evidence, or conflicting claims. It is unnecessary for every question.
| Approach | Use it for | Typical output |
|---|---|---|
| Standard search or chat | A current fact, official link, brief explanation, or narrowly scoped question. | A concise answer based on limited retrieval. |
| Thinking or reasoning mode | Logic, analysis, planning, coding, mathematics, or a problem that may not require extensive web evidence. | A reasoned answer that can still depend on the model’s available knowledge and tools. |
| Deep Research | Market research, literature discovery, policy analysis, product evaluation, technical landscapes, and claims that require evidence from multiple sources. | A longer report with research stages, synthesis, and citations. |
Choose Deep Research when the value comes from investigating the evidence—not merely from receiving a longer answer. A narrowly defined factual question will often be faster and easier to verify with normal search.
How to use Qwen Deep Research
The exact labels can change, but the practical workflow remains consistent: define the question, clarify the scope, let the system investigate, and verify the final report.
1. Open the official Qwen research experience
Go to the official Qwen research page and sign in if the current experience requires an account. Use the official Qwen domain rather than a similarly named third-party site.
If Deep Research is not shown, check the current tool or mode selector, your account status, the Qwen application version, and regional availability. The absence of a control on one account does not prove that the feature has been discontinued globally.
2. Decide how much research the task needs
Qwen’s 2511 announcement introduced two research depths:
- Normal: intended for efficient, broadly useful research.
- Advanced: intended to spend more time on deeper or more demanding investigations.
Use Normal for a bounded report with clear comparison criteria. Choose Advanced, when the option is present, for a multi-part question, conflicting evidence, a longer historical period, several markets, or a topic that requires more source reconciliation. Do not select the deeper mode simply because it sounds more accurate; a precise scope usually improves the result more than an unnecessarily broad investigation.
3. Write a research brief, not a vague question
Replace prompts such as “Research the electric vehicle market” with a brief that defines the decision, geography, period, evidence, comparison dimensions, and expected output.
A useful request should tell Qwen:
- The central research question.
- The person or decision the report should support.
- The countries, markets, organizations, or products in scope.
- The applicable time period.
- The criteria that must be compared.
- The preferred primary or authoritative sources.
- Sources or topics that should be excluded.
- How uncertainty and source disagreement should be handled.
- The desired report structure and level of detail.
4. Upload supporting material only when it is necessary
Qwen announced document and image input for DeepResearch 2511. If the current interface offers uploads, add only material that contributes directly to the investigation—for example, a public report, product specification, survey, or spreadsheet you are authorized to use.
For a PDF source, use our Qwen PDF guide to prepare the file and verify its extracted answers.
Before uploading a file, remove API keys, passwords, personal identifiers, confidential client data, unpublished financial information, medical records, and material you do not have permission to process. File support, formats, limits, and retention should be checked in the current interface and policies rather than inferred from an older announcement.
5. Answer clarification questions carefully
Qwen Deep Research can ask follow-up questions before beginning the full investigation. These questions are useful because they expose ambiguous parts of the request, such as what “best” means, which markets matter, or how current the sources must be.
Use this stage to narrow the task rather than replying “everything” or “all available information.” Specify the decision criteria, acceptable sources, and the evidence needed to support the conclusion.
6. Review the scope before relying on the report
If Qwen displays a research plan, confirm that it covers the right subquestions and excludes irrelevant ones. A plan that misses the core decision will produce a polished but unhelpful report.
- Check that the time period is correct.
- Confirm that the requested countries, companies, or products are included.
- Look for at least one step that searches for evidence against the expected conclusion.
- Confirm that the plan prioritizes primary sources where they exist.
- Remove subtopics that do not help the intended decision.
7. Treat completion as the start of verification
When the report is ready, read the conclusion, but then inspect the evidence supporting its most important claims. A finished status means the generation process ended; it does not certify that every citation, number, inference, or recommendation is correct.
A practical Qwen Deep Research prompt template
The following structure is adaptable to market, technical, product, historical, or policy research:
Research question:
[State one precise question.]
Decision this report should support:
[Explain what the reader must decide or understand.]
Audience:
[Define the intended reader and their technical level.]
Scope:
- Geography:
- Time period:
- Organizations, products, or populations:
- Topics to include:
- Topics to exclude:
Evidence requirements:
- Prioritize original data, official documents, peer-reviewed research,
regulatory sources, company filings, and current technical documentation.
- Distinguish facts, estimates, forecasts, and interpretations.
- Identify evidence that challenges the leading conclusion.
- Explain why credible sources disagree.
- Do not fill evidence gaps with assumptions.
Comparison criteria:
[List the exact dimensions that must be evaluated.]
Output:
- Begin with a concise answer.
- Provide a comparison table where the data supports one.
- Cite the source beside each material claim.
- Include publication dates and applicable geography for statistics.
- Add a section for limitations and unresolved questions.
- Finish with a decision-oriented conclusion.Example: researching software options
Research customer-support platforms suitable for a 40-person B2B software
company selling in Germany and the United Kingdom.
Compare no more than six platforms using information current through
October 2026. Evaluate total pricing structure, EU data-hosting options,
security documentation, Salesforce integration, multilingual support,
AI-assisted ticket handling, implementation effort, and public support
quality evidence.
Prioritize each vendor's official pricing, security, data-processing, and
integration documentation. Use independent sources only for implementation
experience and documented limitations. Do not infer a feature from marketing
language when technical documentation does not confirm it.
Identify evidence both supporting and challenging each shortlist candidate.
Explain material disagreements between sources. Separate verified facts from
estimates and recommendations.
Output:
1. Executive answer
2. Requirements table
3. Evidence-based comparison
4. Three-candidate shortlist
5. Risks and unanswered questions
6. Source-verification appendixThis prompt does not tell the system to prove that one vendor is best. It defines a neutral investigation and requires evidence against the emerging conclusion.
How to verify Qwen Deep Research citations
A report can contain real links and still misrepresent what those pages say. Verify the claims that would materially affect a decision, publication, purchase, academic submission, or high-stakes action.
- Identify the decisive claims. Start with the conclusion, key figures, rankings, dates, legal statements, and causal explanations.
- Open the cited source. Do not rely on the report’s description of it.
- Confirm that the page exists and is the intended document. A link may point to a homepage, summary, repost, or unrelated section.
- Find the supporting passage. Verify that the source actually supports the attached sentence rather than a narrower or different claim.
- Check the date. Distinguish the source’s publication date, update date, and the date on which the reported event occurred.
- Classify the source. Determine whether it is an original record, official documentation, peer-reviewed research, journalism, vendor marketing, opinion, or an AI-generated summary.
- Inspect numerical scope. Confirm the unit, currency, country, sample, period, methodology, and whether the figure is observed, estimated, or forecast.
- Look for omitted qualifications. Read surrounding text, footnotes, eligibility rules, and methodology.
- Cross-check disputed facts. Use another primary or authoritative source when the claim is controversial, surprising, or consequential.
- Revise unsupported conclusions. Mark uncertainty instead of keeping a confident statement that the available evidence does not establish.
Common citation failures
| Failure | What to check |
|---|---|
| Real source, wrong claim | The page exists, but its text does not support the report’s sentence. |
| Partial support | The citation supports one clause but not the larger conclusion attached to it. |
| Stale evidence | An old source is used for a price, law, model, executive role, product feature, or market figure that may have changed. |
| Secondary instead of original | A blog cites a study, filing, or official dataset that should be opened directly. |
| Missing numerical context | A statistic is presented without the applicable location, period, sample, unit, or methodology. |
| Source disagreement hidden | The report selects one figure without explaining different definitions or measurement methods. |
| Inference presented as quotation | The report’s synthesis is written as though the source explicitly reached that conclusion. |
For academic work, cite the original papers, datasets, books, or official records—not the Qwen-generated report as a substitute. Follow the institution’s AI-use and disclosure rules, and verify all bibliographic details directly.
How to evaluate source quality and disagreement
Source quality depends on the claim. A government statistical agency may be authoritative for population data, while a software vendor’s documentation is more appropriate for its current API parameters.
A useful default hierarchy is:
- Original laws, court opinions, standards, regulatory publications, official statistics, filings, and public records.
- Peer-reviewed studies, original experiments, technical reports, and primary datasets.
- Official company documentation, pricing pages, product policies, and release announcements.
- Reputable specialist or investigative reporting.
- Qualified expert analysis that clearly identifies its evidence.
- Community reports and user discussions, used as leads rather than definitive proof.
- Unsourced summaries, promotional pages, and AI-generated pages.
When credible sources disagree, investigate definitions, dates, locations, sample design, revisions, funding, and whether the figures are forecasts or observations. Preserving a material disagreement is more accurate than forcing a single answer that the evidence does not support.
Reports, webpages, podcasts, and sharing
Qwen announced in 2025 that completed Deep Research reports could be transformed into a live webpage and a multi-speaker podcast. Those announcements establish that the capabilities existed in the announced experience; they do not guarantee that every current account, region, or application still exposes the same controls.
After generating a report, check the current interface for options such as Create, Export, Download, Share, Webpage, or Podcast. Before using any output:
- Confirm whether the output is downloaded privately or hosted at a shareable URL.
- Check whether anyone with the link can open it.
- Verify whether citations survive the conversion.
- Distinguish source-derived charts and images from AI-generated illustrations.
- Review quotations and copyrighted material before publishing.
- Remove private material that should not appear in a webpage or audio program.
Transforming a report into a polished webpage or podcast does not correct errors in the original research. Verify the report first, then create derivative formats.
Is Qwen Deep Research free?
Qwen’s public website describes the consumer assistant as free to use, and the original DeepResearch launch invited users to start researching for free. That does not establish unlimited Deep Research usage, identical availability for every account, or a permanent numerical allowance.
| Route | Cost interpretation |
|---|---|
| Consumer Qwen Deep Research | May be available without a separate payment, but the current account interface is the authoritative source for task allowances, resets, queueing, and mode access. |
| Model Studio API | Usage-based paid service. The current official documentation lists no free quota for the mainline Deep Research model. |
| Tongyi DeepResearch | The model and framework are released under Apache-2.0, but operation can require compute, storage, search APIs, page-reading services, parsing tools, and engineering work. |
Do not interpret “open source” or “free to access” as meaning that research has no operational cost. Consumer access, managed API usage, and self-hosting have different cost structures.
Privacy and confidential research
Review the current Qwen Privacy Policy and Qwen Terms of Service before uploading non-public material or sharing a generated report. Do not assume that the consumer product, Model Studio API, third-party search tools, and a self-hosted deployment use identical data policies.
- Remove personal information that is not required for the research question.
- Redact passwords, keys, account identifiers, private URLs, and confidential names.
- Use public documents or files that you are authorized to process.
- Check organizational rules before submitting client, employee, legal, financial, or regulated information.
- Review link permissions before sharing a report, webpage, or podcast.
- Assume that external search, model, or MCP services may receive the information necessary to perform their configured task.
A local model can change where inference happens, but it does not automatically make a research workflow offline. Search APIs, page retrieval, file parsing, external summarizers, package downloads, and other tools can still send data over a network.
Qwen Deep Research API for developers
Alibaba Cloud Model Studio provides a separate managed service named qwen-deep-research. It is not an API wrapper around the consumer Qwen interface, and its configuration should not be copied from a general Qwen API tutorial.
| Item | Current official specification |
|---|---|
| Provider | Alibaba Cloud Model Studio |
| Region | China (Beijing) only; the API key must match the Beijing region. |
| Main model ID | qwen-deep-research |
| Dated snapshot | qwen-deep-research-2025-12-15 |
| Workflow | Two stages: clarification followed by full research. Clarification can be skipped through the documented native parameter when appropriate. |
| Interface | Native DashScope interface through the documented Python SDK or HTTP request. |
| Streaming | Required. |
| OpenAI compatibility | Not supported by the current Deep Research documentation. |
| Java SDK | Not currently supported for this model. |
| Context window | 1,000,000 tokens. |
| Maximum input | 997,952 tokens. |
| Maximum output | 32,768 tokens. |
| Image input | JPEG, PNG, BMP, and WebP; up to five images and 10 MB per image under the current documentation. |
| Mainline price | US$7.742 per 1 million input tokens and US$23.367 per 1 million output tokens. |
| Free quota | None listed for the mainline model. |
| Rate limits | 120 requests per minute and 1,200,000 tokens per minute under the current model page. |
These specifications were verified on October 7, 2026. The official documentation marks pricing for the dated snapshot as undetermined, so it should not be assigned the mainline model’s rate without confirmation.
The API can return either a detailed report or a shorter summary format. Its streamed response can also expose research phases, status events, references, and search-round information. Applications must handle metadata and progress chunks rather than assuming every streamed event contains final report text.
Before implementing it, read the current Qwen-Deep-Research developer guide, native API reference, and model specification and pricing page. Do not combine a key, endpoint, model ID, or SDK pattern from another Alibaba Cloud region or a general OpenAI-compatible Qwen example.
Tongyi DeepResearch and local deployment
Tongyi DeepResearch is a separate open-source model and research framework from Alibaba-NLP. The current project uses the Apache-2.0 license and provides the model Tongyi-DeepResearch-30B-A3B.
- Model scale: 30.5 billion total parameters and approximately 3.3 billion activated per token, according to the technical report.
- Context: the current model materials list a 128K context window.
- Purpose: long-horizon research involving planning, search, reading, and synthesis.
- License: Apache-2.0 for the published project and model card.
Downloading the weights does not create a complete offline research application. The official framework can depend on separate components for web search, academic search, webpage reading, summarization, file parsing, and sandboxed code execution. Its example configuration references services such as Serper, Jina, DashScope, an OpenAI-compatible summarization endpoint, and SandboxFusion.
Self-hosting therefore requires more than model inference. A production deployment needs suitable compute, tool credentials, access controls, monitoring, timeout handling, protection against malicious web content, and a method for validating retrieved sources.
Use the official model card and the repository’s current instructions rather than assuming that the consumer Qwen workflow, Model Studio API, and open-source framework share the same features.
What Qwen Deep Research is suitable for
| Good starting point | Not sufficient by itself |
|---|---|
| Initial market and competitor landscapes | Investment, lending, or trading decisions |
| Finding original sources and relevant terminology | Medical diagnosis or treatment decisions |
| Comparing public product documentation | Legal advice or a final interpretation of law |
| Background briefings on an unfamiliar topic | Safety-critical engineering approval |
| Mapping disagreements between public sources | Compliance certification or security assurance |
| Preparing questions for a specialist | Final academic evidence without checking original sources |
| Organizing a large set of public information | Due diligence that depends on confidential or non-public records |
| Developing hypotheses for further investigation | Claims that require an original experiment, audit, interview, or field study |
The tool is strongest as a structured starting point and source-discovery assistant. It does not replace subject-matter expertise, access to restricted databases, original research, or accountable human review.
Important limitations to check in every report
- Search visibility bias: easily indexed pages may receive more attention than paywalled, regional, archival, or database-only sources.
- Citation mismatch: a cited page can be real while failing to support the attached sentence.
- Freshness: prices, laws, software features, market figures, and public roles can change after a source is published.
- Language and regional bias: the research may favor sources that are easier to retrieve or process.
- Paywall and database limits: inaccessible material can leave an important gap that is not obvious in the final report.
- False reconciliation: conflicting figures may be combined without recognizing that they use different definitions or periods.
- Weak numerical reasoning: derived calculations can be wrong even when the input statistics are cited correctly.
- Transformation loss: a webpage, summary, or podcast may omit caveats or source detail contained in the full report.
Common Qwen Deep Research problems
| Problem | Likely cause | What to do |
|---|---|---|
| Deep Research is missing | Account, region, platform, application version, rollout, or usage eligibility differs. | Use the official Qwen site, sign in, check the current tools or mode menu, update the app if applicable, and review account notices. |
| The task does not start | A clarification question is waiting, the scope is too broad, a usage allowance was reached, or the service returned a temporary failure. | Look for a pending question or notice, narrow the prompt, and retry once rather than submitting many duplicate jobs. |
| The report is too generic | The prompt lacked a decision, geography, period, source requirements, or comparison criteria. | Rewrite the request as a structured research brief and require primary sources and contradictory evidence. |
| Citations do not support the text | The synthesis overextended the source or attached one citation to several claims. | Open each source, identify the unsupported sentence, and request a correction limited to verifiable evidence. |
| An uploaded file is ignored | The format, size, extraction quality, scan quality, or account capability is unsuitable. | Check current upload rules, use a searchable file, reduce its size, and ask Qwen to confirm what it can read before starting the full report. |
| Webpage or podcast options are absent | The capability may not be enabled for the current account, region, platform, or report. | Check the current Create, Export, and Share controls rather than relying on a 2025 announcement or an older tutorial. |
| API authentication fails | The API key is not from China (Beijing), the workspace or endpoint is wrong, or the consumer and Model Studio routes were mixed. | Recreate the configuration using only the current Beijing Model Studio documentation. |
| OpenAI SDK code fails | The current Deep Research API does not support the OpenAI-compatible interface. | Use the documented native DashScope request and streaming workflow. |
| The streamed API response appears empty | The application is receiving progress or metadata events rather than final report text. | Handle the documented phases, statuses, references, and content fields instead of printing one assumed response property. |
Frequently asked questions
Is Qwen Deep Research the same as Qwen Thinking?
No. Thinking is primarily a reasoning mode, while Deep Research is designed for longer investigations involving planning, repeated retrieval, source synthesis, and a cited report. The exact tools available to either mode can change.
Are Qwen Deep Research citations reliable?
They are useful starting points, but their presence does not certify the report. Open each important citation and confirm that it supports the claim, is current enough, and uses the correct geography, period, unit, and methodology.
Can Qwen Deep Research use academic sources?
It can discover publicly accessible academic and research material, but access to paywalled journals, private databases, and institutional subscriptions may be limited. Cite and evaluate the original paper rather than relying on Qwen’s summary.
Can I use a Qwen report for university work?
Use it for source discovery, topic mapping, and question development only within your institution’s rules. Verify the underlying sources, write the analysis yourself, cite the originals, and disclose AI assistance when required.
Is the Qwen Deep Research API OpenAI-compatible?
No. The current Alibaba Cloud documentation says the qwen-deep-research service uses the native DashScope interface and does not support the OpenAI-compatible API.
Is Tongyi DeepResearch the open-source version of the Qwen consumer feature?
Tongyi DeepResearch is an official open-source research model and framework, but it should not be described as an exact open-source copy of the consumer Qwen feature unless Qwen explicitly confirms that relationship. Their interfaces, tools, deployment requirements, and supported outputs differ.
Can Tongyi DeepResearch run completely offline?
The model can be self-hosted, but the complete reference workflow can depend on web search, page retrieval, file parsing, summarization services, and a sandbox. Running the weights locally does not automatically make the full research system offline.
Sources checked
The following primary sources were checked on October 7, 2026:
- Official Qwen research page.
- Official Qwen DeepResearch announcement.
- Qwen Privacy Policy and Qwen Terms of Service.
- Alibaba Cloud Model Studio Qwen-Deep-Research guide.
- Qwen-Deep-Research native API reference.
- Qwen-Deep-Research model specification, pricing, and rate limits.
- Official Alibaba-NLP Tongyi DeepResearch repository.
- Official Tongyi-DeepResearch-30B-A3B model card.