Kimi Deep Research: Access Test, Sources & Limits

Kimi Deep Research is a research agent that plans a task, searches the web, evaluates sources and produces a cited long-form report. It is designed for work that needs more investigation than an ordinary chat answer, such as literature reviews, market research, policy analysis and multi-source fact finding.

This guide separates three things that are often blended together:

what Kimi officially says the product does;
how to use it without losing control of the research scope; and
what our authenticated access attempt observed and how the future performance test will be run.

Test status: On August 4, 2026, we logged in, selected Deep Research with K3 High, entered a fixed five-question official-source prompt and attempted to start it. Kimi displayed: “Too many people are chatting with Kimi right now. Subscribe to enter a dedicated priority queue!” Research did not begin. We therefore report an access-blocked attempt, not an accuracy, speed, source-quality, citation or credit-use result.

Kimi AI Guide is independent and is not affiliated with Moonshot AI. Vendor-published averages below are labeled as such; they are not our measurements.

What Kimi Deep Research is

Kimi’s current Help Center describes Deep Research as an Agent product powered by Kimi-Researcher. It says the system can clarify the question, plan searches, inspect pages, use tools, reason iteratively and assemble a cited report. The official technical report describes a research agent with parallel search, a text browser and code execution.

The product is distinct from ordinary Kimi Search:

Product	Best fit	Typical interaction	What this page evaluates
Kimi Chat with web search	A current fact or a bounded question	Short conversational answer	Not the main subject of this page
Kimi Deep Research	A question requiring planning, many sources and synthesis	Asynchronous research task and long report	End-to-end report coverage, factual support and source handling
Kimi Agent	Creating or modifying a deliverable with multiple tools	Autonomous task execution	Covered separately in the Kimi Agent guide

For a focused benchmark of short search answers, see our Kimi Search Citation Accuracy Test. That test uses many atomic questions; this page examines complete research reports.

Current documented capabilities and limits

The table below records official product statements checked on August 4, 2026. “Vendor-reported” means Kimi published the figure; we have not yet reproduced it.

Item	Current official description	Evidence status
Underlying product model	Kimi-Researcher	Officially documented
Research workflow	Clarification, reasoning, active search, iterative review, tool use and report generation	Officially documented
Typical execution time	10-25 minutes, asynchronous	Vendor-reported range; live measurement pending
Context length	128K tokens	Officially documented; usable task capacity not independently measured
Typical activity	About 23 reasoning steps, 74 planned keywords and 206 discovered URLs, with 3.2% retained	Vendor-reported averages, not fixed quotas
Text report	More than 10,000 words on average, with about 26 traceable sources	Vendor-reported averages, not a promise for every task
Visual report	Interactive, shareable HTML report with structured layout and mind maps	Officially documented; export test pending
Saving results	Text reports can be copied or downloaded as PDF or Word; visual reports can be shared and saved from preview	Officially documented; export test pending
Execution behavior	The task continues asynchronously; Kimi advises refreshing rather than stopping an apparently idle run	Officially documented
Credits	Membership features share a credit pool and consumption depends on tokens; Kimi gives roughly 5-10% of free-tier credits as an example for a deep research report	Vendor estimate; actual consumption varies and will be recorded per run
Simple questions	Kimi recommends standard chat for simple Q&A	Officially documented

These figures describe the consumer product, not a Deep Research API. Kimi’s current API model-selection help says a Deep Research API is not available. Do not reuse the 128K Deep Research context figure as an API model limit; Kimi publishes separate API limits for models such as K3 and K2.6.

How to use Kimi Deep Research

1. Choose a question that needs research

Use Deep Research when the answer depends on several sources, disputed evidence, a timeline or a structured comparison. Use ordinary chat when one authoritative page can answer the question directly.

Weak request:

Tell me about the AI market.

Researchable request:

Using sources published from January 2024 through July 2026, compare the
adoption of generative AI by small businesses in the United States and the
United Kingdom. Separate survey results from estimates, identify the sample
and sponsor of every survey, and do not combine metrics with different
definitions. Prioritize government statistics and original survey reports.

2. Define the evidence rules before the task starts

A strong request states:

the exact question and intended reader;
the geographic and time boundaries;
the fields that must be covered;
preferred and excluded source types;
how conflicting evidence should be handled;
what must be marked not found rather than inferred; and
the required format and length.

Use this reusable prompt frame:

Research question: [one bounded question]
Audience: [who will use the report]
Cutoff: Use information published by [date and time zone].
Scope: [countries, products, years or documents]
Required fields: [list every field]
Source policy: Prioritize [primary/official/peer-reviewed sources].
Do not use: [aggregators, anonymous posts, search snippets, etc.]
Conflict rule: Show conflicting figures side by side; do not average them.
Missing-data rule: Write "not found" and describe the search performed.
Citations: Cite every material factual claim at the point where it appears.
Output: [table/report/appendix], no more than [length].

3. Use the clarification stage

Kimi says Deep Research asks follow-up questions before searching. Treat this as a scope checkpoint. Correct an unwanted geography, date range, definition or deliverable here. A long clarification answer can introduce new ambiguity, so answer with short, explicit constraints.

4. Let the asynchronous task finish

Kimi documents a typical 10-25 minute run and says the task can continue while you leave the conversation. If the page appears unchanged, refresh it rather than clicking Stop output. The official FAQ says manually stopping or closing a running task can still consume credits.

5. Audit the report before using it

Do not judge a report by its length, citation count or design. Check:

whether every requested field was answered;
whether each citation opens;
whether the cited passage supports the nearby claim;
whether a primary source was available but ignored;
whether dates and units match the source;
whether estimates are labeled as estimates;
whether contradictory sources were disclosed; and
whether recommendations go beyond the evidence.

Observed access attempt: August 4, 2026

We performed a signed-in access attempt before assigning any performance claim.

Field	Observed state
Account state	Logged in; subscription tier is not inferred from the blocker
Product mode	Deep Research selected
Displayed model	K3 High
Input	Five-question official-sources-only task scope; the complete verbatim input was not independently retained
Start result	Blocked before research execution
Visible message	“Too many people are chatting with Kimi right now. Subscribe to enter a dedicated priority queue!”
Search activity	None observed; the research workflow did not start
Report or citations	None generated
Performance scoring	Not applicable

The queue message demonstrates the access state for this account at that moment. It does not show that Deep Research failed the questions, that paid access always works, that free access never works, or that the documented 10-25 minute execution range is inaccurate.

Normalized record of the submitted task scope

The priority-queue modal obscured part of the input in the retained screenshot. The block below is a normalized reconstruction of the verified five-question scope and constraints used for the attempt; it is not presented as a complete screenshot-visible verbatim transcript.

Create a concise research report answering these five questions. Use only
official Kimi Help Center, platform.kimi.ai documentation, Moonshot AI
technical pages, or official model cards - never search snippets or
third-party sources. Answer each item in a table with the exact official
value, a short caveat, and the direct source URL per row:

1. What context limit is documented for Kimi Deep Research?
2. What are the documented per-file size and files-per-session limits in
   Kimi Chat?
3. What context windows are documented for the kimi-k3, kimi-k2.7-code and
   kimi-k2.6 API models?
4. Are Project files fully injected into every turn?
5. Do Kimi Membership, Kimi Code and Kimi API share one balance?

Write "not published" when an official source does not publish a requested
value. Keep the report under 300 words and add a limitations section. Record
the access date as August 4, 2026.

The send attempt produced the priority-queue message before a research plan, search query, visited URL or answer appeared. No answer row can therefore be scored.

Kimi Deep Research with K3 High showing a priority-queue subscription message before the five-question research task started. — Authenticated Kimi Deep Research access attempt on August 4, 2026. The priority-queue message appeared before research began; this is access evidence, not a performance result.

Planned independent performance test

The access attempt above is complete. The performance methodology below is still planned and will be executed only when the research task can start. A retry receives a new run ID; it does not overwrite the blocked attempt.

We plan two complementary tasks. DR-01 uses the same five-question official-source audit scope with a known answer key. DR-02 is an external public-data synthesis task. Together they can reveal different failure modes without pretending that two reports establish universal performance.

Fixed environment rules for a future run

For both tasks:

start a new English-language Deep Research conversation;
use the same authenticated account and record the plan without revealing identity;
record the exact product and model labels visible in the interface;
do not upload files or provide source URLs in the prompt;
submit the frozen prompt without adding hints from the answer key;
do not edit, regenerate or steer the report after execution begins;
retain every failure and off-topic source, not only the best output;
capture the start and completion timestamps in UTC;
record credits before and after using the smallest visible unit; and
stop and classify the run separately if login, queue access or sufficient credits are unavailable.

DR-01: Five-question Kimi documentation audit

Question: Can Deep Research return the five requested product facts from the correct official consumer or API scope, preserve caveats and cite a direct source for every row?

The future run will reuse the submitted prompt printed above.

DR-01 ground-truth ledger

This ledger was prepared independently of any model output. Every source and value must be rechecked immediately before a future scored run because Kimi documentation can change.

Question	Ground truth at protocol freeze	Primary official source
Deep Research context	128K tokens	Deep Research FAQ
Kimi Chat file limits	Up to 100 MB per file and up to 50 files per session	Kimi overview
API context windows	`kimi-k3`: 1M tokens; `kimi-k2.7-code`: 256K tokens; `kimi-k2.6`: 256K tokens	Kimi Platform introduction and model list
Project file injection	No. Kimi documents Project files as read on demand rather than fully injected into every turn	Projects
Product balances	The three products do not share one universal balance. Current membership rules include Kimi Code in the membership credit pool, while Open Platform API balance and keys remain separate	Membership credit rules and API troubleshooting

DR-02: External primary-source synthesis

Question: Can the product extract closely related estimates without changing years, units, confidence language or population groups?

Exact prompt:

Using only World Health Organization primary publications available by
23:59 UTC on August 4, 2026, write an evidence brief on the global malaria
burden in 2024 as reported in the World malaria report 2025.

Report: estimated global cases and deaths for 2024; the comparable 2023
figures; change in cases; the African Region's share of cases and deaths;
the share of African-region malaria deaths among children under five; the
number of countries using malaria vaccines in routine programmes; the number
of countries and children reached by seasonal malaria chemoprevention; and
WHO's estimate of cases and deaths averted by wider use of new tools in 2024.

Preserve units, years and words such as "estimated." Distinguish report
estimates from reported surveillance counts. Use the report, executive
summary, annexes or WHO release rather than third-party summaries. Cite every
number at the point of use and include page, table or section when available.
If two WHO pages disagree, show both values and do not silently select one.
Keep the brief between 1,000 and 1,500 words and end with a claim-to-source
appendix.

Fixed clarification response:

Keep the scope global and limited to the requested 2023-2024 comparison.
Do not provide medical advice, country rankings or projections. Prioritize the
World malaria report 2025 and its annexes over later summaries, while listing
any material discrepancy you find.

DR-02 ground-truth ledger

Field	Ground truth at protocol freeze	Primary official source
2024 global cases	282 million, estimated	WHO 2025 report release
2023 global cases	273 million, implied by WHO’s stated increase of about 9 million; verify against report table before scoring	WHO 2025 executive summary
2024 global deaths	610,000, estimated	WHO malaria fact sheet
2023 global deaths	598,000, estimated	WHO malaria fact sheet
African Region share	94% of cases and 95% of deaths in 2024	WHO 2025 executive summary
Children under five	75% of deaths in the African Region	WHO 2025 executive summary
Vaccine programmes	24 countries had introduced malaria vaccines into routine immunization programmes	WHO 2025 report release
Seasonal chemoprevention	20 countries; 54 million children reached in 2024	WHO 2025 report release
Wider-tool estimate	170 million cases and 1 million deaths averted in 2024	WHO 2025 report release

The 2023 cases value must be checked in the full report immediately before scoring. A value inferred by subtraction is not treated as a report quotation.

Scoring rules

We report separate metrics rather than hiding them inside one star rating.

Required-field coverage

Each requested field receives:

2 – present, unambiguous and in the required format;
1 – present but incomplete, ambiguous or outside the requested format;
0 – omitted or replaced by a different fact.

Coverage = points earned / maximum possible points x 100

Claim-level citation audit

Before opening any cited source, split the report into atomic, externally checkable claims. A sentence containing two dates and one quantity can create three claim rows.

Each claim is then scored:

Dimension	2	1	0
Citation completeness	Citation clearly attached to the claim	Citation placement is ambiguous	No citation
Entailment	Source supports the whole claim	Source supports only part	Source does not support it or cannot be located
Source quality	Requested primary source	Credible secondary source where primary was requested	Unsuitable, anonymous or search snippet
Freshness	Current for the stated cutoff and reference period	Date unclear or newer source was available	Stale for the claim
Numeric fidelity	Value, unit, population and period all match	One non-material qualifier is missing	Material mismatch

We also report these counts separately:

fabricated or malformed URLs;
links that do not open for a normal reader;
unsupported quotations;
undisclosed source conflicts;
estimates presented as observed counts; and
claims added outside the requested scope.

No overall product grade

Two reports cannot establish general accuracy or reliability. We publish the coverage and citation measures separately, show the raw rows and describe failures. We do not convert these results into stars, a universal percentage or a claim that Kimi Deep Research is “best.”

Observed result and planned runs

Run	Task	Date UTC	Research started	Duration	Credits used	Sources cited	Coverage	Citation entailment	Status
DR-ACCESS-01	Five-question official-source prompt access attempt	August 4, 2026	No	Not applicable	Not measured	Not applicable	Not scored	Not scored	Access blocked by priority queue before execution
DR-01-R1	Future scored five-question audit	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Planned
DR-02-R1	Future WHO evidence synthesis	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Planned

Not scored is not zero. The blocked attempt produced no research output to evaluate. A future completed run will be added as a new row rather than replacing DR-ACCESS-01.

Raw-data tables

Run metadata

Run ID	Protocol version	Account state	Interface label	Model label shown	Language	Browser/OS	Test region	Start credits	End credits	Export obtained	Evidence bundle
DR-ACCESS-01	1.0 access attempt	Authenticated; tier not asserted	Deep Research	K3 High	English	Browser session; version not recorded	Not independently determined	Not recorded	Not recorded	No	Original priority-queue screenshot shown above
DR-01-R1	1.0	Pending	Deep Research	Pending	English	Pending	Not independently determined	Pending	Pending	Pending	Planned
DR-02-R1	1.0	Pending	Deep Research	Pending	English	Pending	Not independently determined	Pending	Pending	Pending	Planned

Required-field results

Run	Field ID	Required field	Output value	Ground-truth value	Coverage 0-2	Correct	Citation IDs	Notes
DR-ACCESS-01	All	Five requested answers	No output generated	See frozen ledger	Not applicable	Not scored	None	Priority-queue message appeared before research began

Claim-to-source audit

Run	Claim ID	Exact atomic claim	Citation URL	Link opens	Source date	Completeness 0-2	Entailment 0-2	Quality 0-2	Freshness 0-2	Numeric fidelity 0-2/NA	Decision note
DR-ACCESS-01	None	No output claim exists	None	Not applicable	Not applicable	Not applicable	Not applicable	Not applicable	Not applicable	Not applicable	Access-only event

Execution timeline

Run	UTC time	Stage	Visible query, URL or event	Observation	Screenshot ID
DR-ACCESS-01	August 4, 2026; exact UTC time not recorded	Access attempt	Deep Research selected; K3 High displayed; send attempted	Priority-queue subscription message appeared before execution	`kimi-deep-research-priority-queue-2026-08-04.png`

The completed tables should be downloadable as CSV in the Kimi AI Test Lab alongside a checksum or immutable file version.

How to verify a Kimi Deep Research report yourself

Use a small claim ledger rather than reading citations casually:

Claim	Citation opens?	Passage supports it?	Source is primary?	Date fits?	Your decision
Copy one factual claim	Yes/No	Full/Partial/No	Yes/No	Yes/No	Keep/Correct/Remove

Open the destination page, search for the exact number or phrase, inspect the surrounding qualifier and confirm that the page is about the same period and population. A real link can still be a bad citation.

For decisions involving health, law, finance or safety, use the report as a research aid and review the primary material with an appropriately qualified professional.

Practical limitations

The observed result is access-only

The August 4 attempt ended at a queue gate before the research workflow began. It cannot answer whether Kimi would find the correct values, use the requested source scope, finish within the documented time range or consume a particular amount of credit. It also cannot establish general availability: queue conditions and subscription entitlements can change.

One run is a case study, not a reliability rate

Deep Research outputs and live web results can vary. The two planned performance runs can expose concrete successes and failures under dated conditions, but they cannot estimate how often the product will succeed across every subject.

The web changes after the test

A linked page can be revised, moved or removed. We record access dates and retain a permissible excerpt, screenshot or checksum so a later reader can understand what was scored.

Source count is not source quality

Twenty citations can repeat the same press release or fail to support the nearby text. We score claims and source passages, not just the total number of links.

Product labels and plans can vary

Availability, model labels, credits and exports may depend on account, rollout, region and plan. We report only the tested account state and do not generalize it to every user.

The interface does not expose every internal action

Visible queries and URLs are observations from the product interface. They may not constitute a complete internal trace. We describe them as visible activity, not a full chain of reasoning.

Ground truth can also be wrong or ambiguous

Official sources can conflict. When they do, the evaluator records each value and applies the predeclared source hierarchy instead of silently picking the convenient answer.

Frequently asked questions

Did Kimi Deep Research complete our test?

No. We reached the signed-in Deep Research interface, selected K3 High and attempted to submit the fixed prompt, but a priority-queue subscription message appeared before research began. We classify this as access blocked and publish no performance score.

Is Kimi Deep Research free?

Kimi says Deep Research is available with a limited free allowance and that membership features use a shared credit pool. The amount used depends on the task and token consumption; Kimi’s 5-10% figure for a report is explicitly a rough free-tier example, not a fixed price.

How long does Kimi Deep Research take?

The official typical range is 10-25 minutes. That is vendor documentation, not our current measured result. Actual time can depend on the task, service state and account.

How many sources does it use?

Kimi reports about 26 traceable sources on average. An average is neither a minimum nor evidence that every citation is correct. A future completed test will audit each claim-to-source relationship.

Can I leave the page while it runs?

Kimi says the task runs asynchronously and you can return later. Its Help Center recommends refreshing an apparently idle page rather than stopping the task.

Can I use Deep Research through the Kimi API?

Kimi’s current API help says a Deep Research API is not available. The consumer Deep Research product and ordinary Kimi API model calls should be treated as separate products.

Is a cited Kimi report safe to trust without checking?

No. A citation can be stale, partial or unrelated even when it opens. Check important claims against the source, especially for high-stakes decisions.

Official sources

Official sources checked August 4, 2026: