Try Kimi at kimi.com →

Kimi-Researcher
The Agent Behind Deep Research

Kimi-Researcher is the research agent that powers Kimi Deep Research. Ask a question and it plans, searches, reads hundreds of pages, checks its facts and writes a long report with sources, usually in 10–25 minutes. Here's how it works and how to get the best from it.

Powers Deep Research 200+ pages per task Cited reports
How are EV battery priceslikely to change by 2030? KBefore I start:Which chemistries, and whichregions matter most to you? LFP and NMC, global view Planning Searching in parallel Reading and checking sources Running numbers in code Writing the report REASONING STEPS 1481216192123 23 on average per task PAGES VISITED 3184680115150182200+ over 200 URLs per task REPORT EV Battery Coststo 2030 1 2 3 sources Visualreport ✓ 10,000+ words, cited
OVERVIEW

Kimi-Researcher at a glance

23Reasoning steps

On average per task, planned by the agent itself.

200+Pages visited

Searched and read for each research task.

10–25Minutes

Typical time for one Deep Research report.

26.9%Humanity's Last Exam

A top score at launch in June 2025.

Model or feature? Kimi-Researcher is the agent model. Deep Research is the button you press in Kimi to use it. Kimi's help centre says Deep Research is powered by Kimi-Researcher.
SPECS

Kimi-Researcher specifications

SpecKimi-Researcher
Launched20 June 2025 (gradual rollout in Kimi)
DeveloperMoonshot AI
What it isAn autonomous research agent model, built on an internal Kimi k-series model
TrainingEnd-to-end agentic reinforcement learning on whole research runs
ToolsParallel real-time search, a text-based web browser and code execution
Typical runAbout 23 reasoning steps and over 200 web pages per task
Where you use itKimi Deep Research on kimi.com and in the Kimi apps
OutputA 10,000+ word Markdown report with sources, plus an interactive visual report
Time per reportUsually 10–25 minutes
CostAgent credits: about 5–10% of a Plus plan's monthly credits per report
APINot available as an API

From Moonshot's Kimi-Researcher launch post and the Kimi Help Center's Deep Research and credits pages. Moonshot hasn't published the model's size.

HOW

How Kimi-Researcher was trained

Instead of following a fixed script, Kimi-Researcher learned the whole research process by trial and reward.

End-to-end learning

One model runs the whole task, from planning to the final answer. It tries many strategies, gets rewarded for correct answers and learns from the full run, not from isolated examples.

Three tools

A parallel real-time search tool, a text-based browser for interactive web pages, and a code tool for calculations and data work.

Hard practice tasks

Training mixed tasks that force good tool use with maths, coding and very hard search questions that need many rounds of searching and checking.

Rewards for efficiency

Correct answers reached in fewer steps earned more, so it learned to research efficiently rather than wander.

Memory management

It learned to keep what matters and drop the rest, so it can run long tasks with dozens of steps without running out of context.

Habits that emerged

Moonshot saw it cross-check conflicting sources and verify even simple-looking answers before replying, without being told to.

RESULTS

Kimi-Researcher results

TestResultNote
Humanity's Last Exam26.9% pass@1, 40.17% pass@4Up from 8.6% before RL training
xbench-DeepSearch69% pass@1Average of 4 runs; ahead of o3 with search tools
FRAMES, Seal-0Strong resultsMulti-step search and reasoning tests
SimpleQAStrong resultsShort factual questions

Moonshot-reported, tested 17–18 June 2025. Search-based results can vary as search engines change.

Why it mattered: almost all of the jump from 8.6% to 26.9% on Humanity's Last Exam came from reinforcement learning on research runs, which showed that agents can learn research skills by practising them end to end.
HUMANITY’S LAST EXAM, PASS@1June 2025 · higher is better Kimi-Researcher during RL training 8.6% at start 26.9% +18.3 points, almost entirely from end-to-end RL Compared with other systems at the time Claude 4 Opus 10.7% Gemini 2.5 Pro 21.6% OpenAI Deep Research 26.6% Kimi-Researcher 26.9% From Moonshot’s Kimi-Researcher launch, June 2025. 2026 models score much higher on this test.

Reinforcement learning took Kimi-Researcher past the leading research agents of mid-2025.

USE

How to use Kimi Deep Research

  1. 1

    Open Deep Research

    Sign in at kimi.com, then choose Deep Research in the sidebar or switch to Deep Research mode in the app.

  2. 2

    Ask a detailed question

    Say what you want to know, who it's for, the time period and the sources you trust.

  3. 3

    Answer its questions

    Kimi-Researcher asks clarifying questions first. Your answers set the direction, so take a moment over them.

  4. 4

    Watch it work

    You can follow the search terms, reasoning and pages it visits as it goes. It usually takes 10–25 minutes.

  5. 5

    Check and follow up

    Read the report, open key sources, then ask follow-up questions in the same chat.

What you get

  • 1
    A full report in Markdown, often 10,000+ words, with a table of contents and linked sources.
  • 2
    A visual report as an interactive web page you can share.
  • 3
    Credits back automatically if the task fails because of a tool error.

Credits and plans: Kimi pricing and is Kimi AI free? Researching your own files instead? See how to analyze PDFs.

PROMPTS

6 prompts for Kimi Deep Research

Copy a prompt into Deep Research and replace the [brackets]. Detailed questions get better reports.

Market

Market overview

Research the [market] in [region] from [year] to today: size, growth, main players, pricing trends and what experts expect to [year]. Prefer official statistics and company filings, and flag any figures that sources disagree on.

Academic

Literature review

Review the research since [year] on [topic]. Group studies by finding, note sample sizes and methods, explain where results disagree and list the open questions. Link every study.

Compare

Compare options

Compare [option A], [option B] and [option C] for [use case]. Cover cost, features, risks and who each suits, with a comparison table and a clear recommendation for [my situation].

Policy

Policy explainer

Explain how [policy or law] works in [country], who it affects, recent changes and the main arguments for and against, using official and news sources from the last [12 months].

Company

Company deep dive

Research [company]: business model, revenue trends, leadership, competitors, recent news and main risks. Use filings and reputable news, and note the date of each figure.

Due diligence

Fact-check a claim

Investigate the claim that [claim]. Find the original source, check it against independent evidence, and give a verdict with your confidence level and the strongest counter-evidence.

More ideas in the Kimi prompt library.

LIMITS

Kimi-Researcher limitations

A 2025 model

Its benchmark results date from June 2025. Newer models score much higher on the same tests, and Moonshot hasn't published newer Kimi-Researcher scores.

It can still be wrong

Reports cite sources, but sources can be outdated or misread. Open the citations behind any number you rely on.

Credits go quickly

One report uses about 5–10% of a Plus plan's monthly credits, so free and low-tier plans only get a few a month.

No API or open weights

Moonshot said it planned to open-source the models behind Kimi-Researcher, but we haven't found a public release. It's only available inside Kimi.

FAQ

Kimi-Researcher FAQ

Kimi-Researcher is an autonomous research agent from Moonshot AI, launched in June 2025. It plans, searches the web, reads pages, runs code and writes a cited report on its own. It was trained with end-to-end reinforcement learning and powers Kimi Deep Research.

They're closely linked. Kimi-Researcher is the agent model; Deep Research is the feature in the Kimi app that you use, and Kimi's help centre says it's powered by Kimi-Researcher.

Usually 10–25 minutes. It runs in the background, so you can leave the page while it works and come back to the finished report.

Kimi offers limited Deep Research use on the free plan. Each report uses agent credits, about 5–10% of a Plus plan's monthly allowance, so paid plans get more reports.

A long Markdown report, often 10,000+ words, with a table of contents and linked sources, plus an interactive visual report you can share.

At launch it scored 26.9% on Humanity's Last Exam, slightly ahead of OpenAI Deep Research's 26.6% in Moonshot's comparison, and 69% on xbench-DeepSearch. Those were strong June 2025 results; newer AI models score higher today.

No. Kimi-Researcher is only available through Deep Research in the Kimi apps. The Kimi API offers models such as K3.1 Pro, which you can combine with your own search tools.

Give detailed questions with the audience, time period and sources you trust, answer the clarifying questions carefully, and ask follow-up questions afterwards. Always check the sources behind key numbers.

Start a Deep Research report

Ask a detailed question in Kimi Deep Research and get a cited report in about 10–25 minutes.