DeepSeek, Doubao, Kimi or Qwen: which to use for a paper
The four models write similar paragraphs because they are built for everyone's questions. Academic writing has its own conventions: what reviewers check, how a grant proposal argues, which fields of a reference to verify. A general chatbot guesses from one prompt; a skill written for academic work fixes the method.
Product details checked against each vendor's website and help center on 22 October 2026. Numbers in brackets link to the sources at the end of the page.
Choose by the job in front of you
| What you need to do | Start with | Why |
|---|---|---|
| Rewrite a few paragraphs or explain a concept | Any of them | For one question and one answer the models differ little; use the one you already have open. |
| Return to the same papers and end with a Word file | Kimi | Projects hold 50 files that stay for next time; Docs produces .docx directly.[6][8] |
| Get a first outline or proposal in your university's format | Qwen | Its AI writing tool has a paper form for thesis, course or journal papers with level, subject and length, plus proposal and outline templates.[10] |
| Process hundreds of passages in bulk, or plug into your own scripts | DeepSeek | The API charges ¥4–8 per million output tokens with a 1-million-token context; half price outside weekday 9–12 and 14–18 Beijing time.[16] |
| Polish to Nature journal standards or translate into English | Acadwrite | The Nature polishing skill diagnoses section roles and paragraph logic before editing sentences; it comes from nature-skills, with 50,000 stars. |
| Simulate peer review before submission, then answer reviewers | Acadwrite | Three independent simulated reviewers anchor comments to the text; the response skill tracks each comment and checks for claimed changes that were never made. |
| Check whether AI-supplied references are real | Acadwrite | Checks authors, year, volume, pages and DOI against authoritative bibliographic data, and never swaps in a similar paper when one cannot be found. |
| Write an NSFC, NSF or NIH grant proposal | Acadwrite | The NSFC skill covers the scientific question, rationale, innovation and mock review; a separate skill handles NSF and NIH proposals. |
| Read a whole thesis plus dozens of papers at once | Acadwrite | A 1-million-token context, about eight times Kimi's default; papers come from scholarly databases and citations open the original.[5] |
The jobs in a paper, one by one
Grey text means the vendor's pages do not say. That does not mean the feature is missing, so check in the product.
| DeepSeek, Doubao, Kimi, Qwen and Acadwrite compared on features for writing papers | DeepSeek | Doubao | Kimi | Qwen | Acadwrite |
|---|---|---|---|---|---|
| Where your sources live | Help center describes chats only; re-upload each time[1] | Cloud drive and projects; chats in a project share context[3][4] | Projects: 100 MB per file, up to 50 files; free plan 2 projects, 500 MB[6] | Knowledge base for documents, web pages and notes[11] | Workspace: PDFs, notes and drafts together, ready next time |
| How much it reads at once | Not stated | Not stated | About 128K tokens by default (50,000–60,000 Chinese characters)[5] | 10 files per upload; over 200 pages fails to parse[12] | 1 million tokens |
| Where the draft lives | In the chat; only full chat history can be exported[2] | Writing assistant creates files saved to the cloud drive[3] | Docs creates Word or PDF up to about 10,000 characters, edited locally[7][8] | AI writing; help center does not say whether Word export exists | The draft is a workspace file the AI edits in place; exports to Word |
| Where sources come from | Web search[9] | Web search; scholarly search not stated | Web search, plus an optional academic database with undisclosed coverage[13] | Literature review template; whether it limits to scholarly databases is not stated[10] | Journal articles, arXiv and PubMed, in Chinese and English |
| Citations open the original | Not stated | Not stated | Search answers link sources; deep research citations highlight the original text[13][14] | Shows sources; whether they open the original is not stated[15] | Every citation links to the paper; the reference list is built with the text |
| Academic features | Not stated | Not stated | Docs includes a sample for writing SCI papers[7] | Paper form: type, level, subject, length (8,000 characters by default); proposal and outline templates[10] | 23 academic skills, including Nature polishing and review, reference checking and NSFC proposals |
23 academic skills from GitHub's most popular research skill libraries
Each skill turns one academic task into rules: what a reviewer checks first, which layer polishing starts with, which fields of a reference get verified. The skills come from open-source projects; enable the ones you want in the skill market and they are available in workspace chats.
- 54K
stars for K-Dense scientific-agent-skills, source of 5 skills including hypothesis generation, grant proposals and scientific writing
- 50K
stars for nature-skills, source of 10 skills including Nature polishing, simulated review and reference checking
- 20K
stars for Humanizer-zh, source of the Chinese de-templating skill
- 14K
stars for AI-Research-SKILLs, source of the conference paper and research ideation skills
Star counts as shown on GitHub on 22 October 2026.
Review and verification
Nature simulated reviewused in paper review
Three independent reviewers at Nature standards assess originality, significance and technical soundness, with major and minor comments anchored to the text.
nature-skillsReference checkingused in reference checks
Verifies authors, title, year, volume, pages and DOI against authoritative bibliographic metadata, catching year conflicts and mismatched DOIs.
nature-skillsStatistical reporting review
Checks experimental units, multiple comparisons, effect sizes and uncertainty, and the n, error bars and asterisks in figure legends.
nature-skillsCritical appraisal of claims
Examines design, bias and confounding and statistical misuse, grading evidence with GRADE and Cochrane risk-of-bias frameworks.
K-Dense scientific-agent-skills
Writing and grants
Nature paper writing
Combines rules for paper type, section and target journal to draft the abstract through discussion from the author's own evidence.
nature-skillsResponse to reviewers
Turns reviews into a numbered tracking table, answers each reviewer point by point, drafts the cover letter and checks for invented changes.
nature-skillsNSFC proposals
For Young Scientists, General and Key programs: the scientific question, rationale, research plan, innovation, mock review and pre-submission checks.
nsfc-skillsConference papers
For NeurIPS, ICML, ICLR, ACL and others: the narrative, page allocation by section and what reviewers score.
AI-Research-SKILLs
Polishing and de-AI
Nature polishingused in polishing
Diagnoses paper type, section roles and paragraph logic before sentence-level edits; supports Chinese-to-English and Nature-family journal style.
nature-skillsThesis polishing
For Chinese science and engineering theses: diagnoses or rewrites against more than 60 rules on wording, sentences, punctuation and equations.
graduate-thesis-polish-and-write-skillChinese de-templating
Removes filler and stock patterns against 31 checkpoints while keeping facts, hedges and the author's voice; makes no promise about AI detectors.
Humanizer-zhSentence-level editing
Five passes based on Sainani's Writing in the Sciences: clutter, passive voice, sentence structure, consistent terms, numbers and citations.
sciwrite
Ideas and close reading
Testable hypotheses
Lists competing explanations for an observation and derives the predictions, controls and falsification conditions that separate them, ending in a pre-registered plan.
K-Dense scientific-agent-skillsPaper X-ray
Reconstructs the starting point and the authors' bet, separates load-bearing design from decoration, works key equations with real numbers and checks how far the evidence goes.
nature-skillsSocial science topic design
Eight steps from interest to research space, separating under-studied questions from genuinely open ones and judging doctoral feasibility.
nature-skillsResearch ideation
Ten ideation lenses with a diverge, converge and refine process for finding CS and AI research topics.
AI-Research-SKILLs
The built-in workflows have rules too
Literature reviews, introductions, citation finding and polishing are Acadwrite's built-in workflows, and each specifies what the draft must look like. Ask a general chatbot to write an introduction and you get the most common version: a paragraph each on background, current research and significance, citing the papers it happens to remember.
Literature review
Searches from 4–8 angles and builds a list of 40–60 papers; every outline section states which papers it cites and what it argues; the text cites at least 30. You confirm the list and the outline before it continues.
AI literature reviewIntroduction
Organises 3–5 points as background, progress, gap, this study and significance. Every heading must be a specific claim, never a label like "Background"; background and progress points get 8–12 papers each.
AI introduction writingCitation finder
Cites only sentences that report others' work, data or the state of the field, never the author's own methods, findings or common knowledge. Each candidate is checked against the original; weak matches are flagged.
Citation finderPolishing
Adds no facts or data the original lacks; keeps the argument, terms, data and direction of conclusions; when sentences move, citation numbers move with them and are never renumbered.
AI paper polishing
FAQ
These products change quickly. Check each vendor's help center before relying on a feature.
Is DeepSeek good for writing a paper?
It is fine for rewriting paragraphs and explaining concepts. But its help center describes chats only: no cross-chat library, so sources are re-uploaded every time; no stated Word export; and web search covers web pages, not scholarly databases. It suits single paragraphs, not managing a whole paper.
Doubao or DeepSeek for writing papers?
Their paragraphs are much alike; the products differ. Doubao has a cloud drive and projects: generated files are saved to the drive, and chats in one project share context so you do not repeat background or re-upload files. DeepSeek's help center describes chats only.
How long a paper can Kimi read?
About 128K tokens by default, roughly 50,000–60,000 Chinese characters, enough for a master's thesis of around 30,000. Projects take files up to 100 MB, at most 50. Word files generated with Docs run to about 10,000 characters.
Can Qwen write a thesis directly?
Qwen's AI writing tool has a paper form: choose the type, level, subject and length (8,000 characters by default) and it generates text. The form sets topic and length, not the structure of the argument, and Qwen does not say whether sources come only from scholarly databases, so check every reference.
How is Acadwrite different from these four?
Acadwrite is a workspace for writing papers: sources, notes and drafts together, with a 1-million-token context. Papers come from journals, arXiv and PubMed and every citation opens the original. It also offers 23 academic skills from open-source projects such as nature-skills (50K stars) and K-Dense scientific-agent-skills (54K stars), covering Nature polishing and simulated review, reference checking and NSFC proposals.
Sources
All pages accessed on 22 October 2026. Items marked as seen in the interface were visible on the vendor's website without signing in.
- [1]DeepSeek：Help center: getting started
- [2]DeepSeek：Help center: exporting chat history
- [3]Doubao：Privacy policy 1.4.2 (writing files saved to the cloud drive)
- [4]Doubao：Doubao Work help: projects
- [5]Kimi：Help center: capabilities
- [6]Kimi：Help center: projects
- [7]Kimi：Docs feature page
- [8]Kimi：Help center: Docs and Sheets
- [9]DeepSeek：API docs: change log (web search)
- [10]Qwen：AI writing and research entries (seen in the interface)
- [11]Qwen：Privacy policy (knowledge base)
- [12]Qwen：Help center: file parsing failures
- [13]Kimi：Help center: web search and professional databases
- [14]Kimi：Help center: deep research
- [15]Qwen：User agreement (reference sources)
- [16]DeepSeek：API docs: models and pricing