Kimi / Hands-on research guide
Kimi Deep Research: run a report and check its sources
To use Kimi Deep Research, open Kimi’s official Deep Research page, ask a bounded question, then check the primary source behind each important claim in the finished report. In our one-task test, Kimi got the main dates right but attached some citations that did not support the nearby detail, and its final verification note contradicted a dated NASA announcement. The practical lesson: follow the citation to the exact passage before relying on the answer.
ChinaAI is an independent publication, not Moonshot AI. This guide covers Deep Research in the consumer Kimi app, not the Kimi developer API or a specific underlying model checkpoint. Our first result is an example to audit, not a measured accuracy rate.
By China AI Chat Team · Tested and source-checked 23 September 2026 · Our research method and corrections
How to use Kimi Deep Research
Kimi describes Deep Research as an agent that can clarify a task, search the web and produce a cited report. Its help page describes text and visual deliverables, but features and access may differ by account or task. The consumer research workflow is separate from ordinary Kimi chat.
- Set one research objective. State the period, geography if relevant, the exact questions to answer, preferred primary sources and what to exclude. Ask for a link beside each material claim.
- Choose Deep Research and review the account notice. In the official app, sign in if needed, confirm the research mode and check the allowance shown to your account before submitting.
- Answer any clarification narrowly. If Kimi asks a follow-up, restate the boundaries rather than opening the topic to everything.
- Wait for the same task to finish. Kimi’s FAQ says it can run in the background. Check its status before starting a duplicate request. Once it finishes, save the output and audit the citations before sharing it.
For a single quick answer, standard chat is usually a simpler starting point. Deep Research is more useful when you need a structured, multi-source report and have time to check it.
A prompt designed for verification
We selected historical NASA milestones because original dated announcements let us check whether a link supported a sentence. The source list and pass/partial/fail checks were recorded before sending the following prompt. Kimi did not receive our answer key.
Build a source-checked chronology of the James Webb Space Telescope from December 2021 through July 2022. Cover four milestones: launch, arrival near Sun–Earth L2, completion of optical mirror alignment, and the first public release of full-color images. For each milestone, give the date, what actually happened, and a clickable citation to a primary NASA or ESA source. Distinguish the first image unveiling from the wider image and data release if they occurred on different dates. Explain in one sentence why mirror alignment did not mean that all instrument commissioning was finished. Keep the text report concise (about 600–900 words), avoid later discoveries and speculation, and flag any fact you cannot verify from a primary source.
For another subject, keep the structure: define scope, ask for primary records, specify the event distinctions that matter, and require an explicit uncertainty flag. Do not ask a research assistant to make a weak source look certain.
What happened in our first attempt
On 23 September 2026, we submitted that prompt once through a signed-in Kimi Deep Research account. The account showed the Moderato plan and a K3 High selector; it did not expose the exact checkpoint serving this research task. Kimi asked no clarification question. The task finished in roughly two minutes and returned one Markdown text report, downloaded inside an “All files” ZIP. We did not regenerate the response. The report contained nine distinct NASA/ESA URLs and covered the requested milestones. We did not observe or test a visual report, PDF export or Word export in this attempt.
The report correctly put the launch on 25 December 2021, the L2 arrival burn on 24 January 2022, and completed optical alignment on 28 April 2022. It also separated the 11 July first-image unveiling from the 12 July wider release. These correct milestones did not make every detail in the report safe to reuse.
Our citation audit: where the links did and did not help
We opened the pages Kimi cited and matched them to the wording immediately beside each link. The short quoted phrases below come from the unchanged first report; the table identifies the sources Kimi placed beside those claims. Download the first report, exact prompt and claim audit (ZIP, 9 KB) to check the findings yourself. This September 26 supplement preserves the original Markdown bytes and recorded hash. The report is model-generated test output with known errors; read it alongside the audit.
| Report section | Original claim and source check | Verdict |
|---|---|---|
| Launch | Kimi called the vehicle and site ESA’s contribution, citing a NASA launch release. It supports the date, Ariane 5 and French Guiana, but not that extra wording. ESA describes its contribution as the launcher and launch services at the spaceport. | Partly supported |
| L2 arrival | Kimi wrote into its planned halo orbit, citing NASA’s 24 January note. NASA says the burn inserted Webb toward its final orbit. The sunshield explanation in the same paragraph is absent there, though another NASA note supports it. | Partly supported |
| Optical alignment | The NASA page Kimi cited supports completed mirror alignment and says instrument commissioning still remained. Kimi preserved this distinction. | Supported |
| First images | Kimi’s JPL link for 11 July and NASA link for 12 July support the two release dates. The adjacent links do not support its simultaneous NASA-site posting detail and 10:30 a.m. EDT start time, respectively. NASA’s separate coverage advisory corroborates both. | Dates supported; citation placement incomplete |
| Final verification note | Kimi said the formal end of commissioning was not pinned to a single dated primary announcement. NASA announced on 11 July that commissioning was complete and science would officially begin on 12 July. | Incorrect |
Decision: this one report is useful as a chronology draft, but correct the ESA wording, L2 wording and final commissioning note before citing it elsewhere. Add the missing nearby sources for the scheduling details. One result cannot establish how often Kimi makes these mistakes on other research questions.
How to check your own Deep Research report
- Mark the claims that matter. Dates, amounts, quotations, legal requirements and recommendations deserve particular attention.
- Open each adjacent link. Find the exact passage that supports the adjacent sentence. A relevant-looking article about the same project is not enough.
- Check what event the date names. An unveiling, technical completion and full release may happen on different days.
- Look for what the report says it could not find. Search official archives for a missing primary announcement before accepting an “unverified” note. Our NASA example shows why.
- Keep the corrected record. Mark each claim supported, partly supported, unsupported or untested; preserve the original output and record the source and date of your correction.
If the report affects a consequential decision, verify the original documents yourself and obtain qualified human review where needed. This one historical test did not examine current market, financial, medical or legal research.
Credits, downloads and limits
Kimi’s FAQ, checked 23 September 2026, says Deep Research uses a shared membership-credit pool and describes background execution, downloadable text and visual reports. In our account, the usage display moved from 44.47% to 45.45%, and the task appeared as 0.98% of that account’s subscription credits. Extra Usage was disabled. This is an observed deduction for one task on one plan, not a dollar price or a guaranteed cost for your account. Check your own allowance before submitting. Our Kimi pricing guide explains consumer membership separately from developer API billing.
The downloaded text was a Markdown file in a ZIP. We did not verify the visual report or the PDF/Word paths described by Kimi’s help; their availability should be checked in the current interface. Before sharing any report, inspect both its facts and whether its link or file reveals material you meant to keep private.
Provider documentation and all cited NASA/ESA pages were reviewed on 23 September 2026. The unchanged first report, exact prompt, predeclared protocol and claim audit are retained in our project records. Contact China AI Chat Team with a source-backed correction. Continue with the Kimi overview for other tasks or read our testing method.