Qwen Deep Research: AI Research Agent
Qwen Deep Research is a multi-step AI research agent built into Qwen Chat. Give it a question, and it autonomously searches the web, reads dozens of sources, reasons through the information, and delivers a structured report with citations — all for free. Since its November 2025 update, it also generates interactive web pages from research results (via Qwen3-Coder) and multi-speaker podcasts (via Qwen3-TTS). Under the hood, it runs on a dedicated 30.5B-parameter model called Tongyi DeepResearch.
Navigate this guide:
- What Is Qwen Deep Research?
- The Model Behind It: Tongyi DeepResearch
- Key Features
- Normal vs Advanced Mode
- Output Formats
- Benchmarks
- How to Use It
- Use Cases
- vs ChatGPT Deep Research vs Gemini
- Limitations
- FAQ
What Is Qwen Deep Research?
Qwen Deep Research is an agentic research feature — not just a chatbot that answers questions, but an autonomous agent that plans a research strategy, executes multi-step web searches, evaluates and cross-references sources, and compiles findings into a professional report. It was first launched in early 2025 and has received major updates in October and November 2025.
Unlike standard web search in chatbots (which searches once and summarizes), Deep Research performs iterative, depth-first research: it reads a page, identifies follow-up questions, searches again, reads more sources, and repeats until it has comprehensive coverage of the topic.
The Model Behind It: Tongyi DeepResearch
Deep Research runs on a dedicated model called Tongyi DeepResearch, built specifically for multi-step research workflows:
| Specification | Tongyi DeepResearch |
|---|---|
| Total parameters | 30.5 billion |
| Active parameters | 3.3 billion (MoE) |
| Base model | Qwen3-30B-A3B-Base |
| Architecture | Multi-agent collaborative with decoupled planning/execution |
| Training | RL-optimized for search, evaluation, and synthesis tasks |
The model uses a decoupled architecture where the planning agent and the writing agent operate independently — the planner decides what to research next, while the writer synthesizes findings in parallel. This enables faster report generation without sacrificing depth.
Key Features
Multi-Source Synthesis
Deep Research doesn't just pull from one search engine. It crawls academic databases, news outlets, tech blogs, government datasets, and specialized sources, then triangulates information across them. Every claim in the final report includes a source URL for verification.
File Upload Support
As of the November 2025 update, you can upload PDFs, documents, and data files as reference material. Deep Research will incorporate your uploaded content alongside web sources.
Report Format Control
Specify how you want the output: academic paper style, executive briefing, technical documentation, bullet-point summary, or custom templates. The agent adjusts structure, tone, and depth accordingly.
Interactive Web Pages
A unique feature added in October 2025: Deep Research can generate an interactive web page from its findings using Qwen3-Coder as a backend. This creates a navigable, shareable research artifact beyond a static PDF.
Multi-Speaker Podcasts
Also added in October 2025: the agent can convert research results into a multi-speaker podcast using Qwen3-TTS, generating a natural-sounding audio discussion of the findings.
Normal vs Advanced Mode
The November 2025 (v2511) update introduced dual modes:
| Mode | Search Depth | Report Length | Speed | Best For |
|---|---|---|---|---|
| Normal | 10-20 sources | 5-10 pages | 5-8 minutes | Quick overviews, preliminary research |
| Advanced | 40-80+ sources | 15-30+ pages | 10-20 minutes | Deep dives, academic reviews, competitive intelligence |
Output Formats
- Web preview — read the report directly in Qwen Chat
- PDF export — professionally formatted with table of contents and citations
- Interactive web page — navigable HTML with charts and sections (via Qwen3-Coder)
- Audio podcast — multi-speaker discussion format (via Qwen3-TTS)
- Markdown/HTML — copy raw content for your CMS or documentation
Benchmarks
Tongyi DeepResearch performs competitively on agentic research benchmarks:
| Benchmark | Tongyi DeepResearch | OpenAI o3 | DeepSeek-V3.1 |
|---|---|---|---|
| Humanity's Last Exam | 32.9 | 26.3 | 24.1 |
| BrowseComp | 43.4 | 49.7 | N/A |
| GAIA | 70.9 | 67.4 | 58.2 |
| FRAMES | 90.6 | 82.1 | 75.3 |
Notably, it outperforms OpenAI o3 on Humanity's Last Exam (a challenging multi-domain benchmark) and GAIA (an agentic benchmark requiring tool use and multi-step reasoning).
How to Use It
- Go to chat.qwen.ai and click Deep Research
- Type your research question — be specific about scope, angle, and desired depth
- Choose Normal or Advanced mode
- Optionally upload reference documents (PDFs, data files)
- Review the AI's research plan — add must-cover topics or tweak direction
- Let it run and watch the live feed as it searches, reads, and synthesizes
- Download as PDF, generate a web page, or create a podcast from the results
Prompt Tips
- Use the SCOPE → FORMAT → CONSTRAINTS formula: "Analyze [topic] focusing on [angle]. Format as [type]. Limit to sources from [timeframe]."
- Follow up with drill-down prompts: "Expand section 3 with data from 2025-2026 only"
- Specify citation style if needed: "Use APA format for all citations"
Use Cases
| Sector | Task | Mode |
|---|---|---|
| Academia | Literature reviews with APA citations | Advanced |
| Finance | Quarterly competitive intelligence decks | Advanced |
| Legal | Multi-jurisdiction case law mapping | Advanced |
| Marketing | Consumer sentiment analysis across forums | Normal |
| R&D | Patent landscape scans | Advanced |
| Content | SEO-optimized research articles | Normal |
| Consulting | Market sizing and opportunity analysis | Advanced |
Qwen Deep Research vs Competitors
| Feature | Qwen Deep Research | ChatGPT Deep Research | Gemini Deep Research | Perplexity Pro |
|---|---|---|---|---|
| Price | Free | $20/mo (Plus) or $200/mo (Pro) | Free (limited) | $20/mo |
| Report length | Up to 30+ pages | 10-15 pages | 5-10 pages | Short summaries |
| Sources per report | 40-80+ | 20-40 | 15-30 | 5-15 |
| File uploads | Yes | Yes | Yes | No |
| Web page generation | Yes | No | No | No |
| Podcast generation | Yes | No | No | No |
| PDF export | Yes | Yes | Yes | No |
| Dual modes | Normal + Advanced | Single | Single | Single |
The standout advantage is the combination of free unlimited use plus unique output formats (web pages and podcasts) that no competitor offers.
Limitations
- Source bias — may over-index on Chinese-language sources for some topics; specify "English-language sources only" when needed
- Recency — while it searches in real-time, some specialized databases may have indexing delays
- Verification required — like all AI research tools, critical facts should be verified against primary sources
- No API access yet — currently only available through the Qwen Chat web interface
- Speed variability — Advanced mode can take 15-20 minutes depending on topic complexity and search depth
FAQ
Is Qwen Deep Research really free?
Yes. No subscription, no credit system, no daily limits. Alibaba Cloud offers it as part of the free Qwen Chat platform.
Does it support PDF uploads?
Yes, as of the November 2025 update. You can upload PDFs, documents, and data files as reference material alongside your research query.
Can I use it via API?
Not yet. Deep Research is currently available only through the Qwen Chat web interface. An API endpoint may be released in the future.
How does it compare to just asking Qwen 3.5 a question?
Qwen 3.5 answers from its training data (with optional web search). Deep Research is an agent that performs systematic, multi-step research across dozens of sources. Use Qwen 3.5 for quick answers; use Deep Research when you need comprehensive, cited analysis.
What's the model behind it?
Tongyi DeepResearch — a 30.5B-parameter MoE model (3.3B active) based on Qwen3-30B-A3B-Base, specifically trained for research planning, source evaluation, and report synthesis.