Kimi Deep Research

Kimi Deep Research is the autonomous research mode of Kimi, the assistant built by the Chinese lab Moonshot AI. Released as Kimi-Researcher in June 2025, it takes a question, plans a line of investigation, runs dozens of searches and page reads against it, and returns a structured report with its sources attached rather than a single-shot answer. It is a general-purpose research agent rather than a scientific one: the corpus it works over is the open web, not a curated bibliographic index.
Why it matters
The work it replaces is the first pass over an unfamiliar topic, where the cost is not reading any one source but knowing which sources exist and how they disagree.
It decomposes a question into sub-questions on its own and follows the results where they lead, so the output covers ground the person asking would not have thought to specify.
Reports come back sectioned and cited, which makes them usable as a starting brief for a review or a market scan rather than as prose to be rewritten.
It is free to use on kimi.com, which puts agentic deep research in reach of labs without a tooling budget.
Key features & advantages
Multi-step agentic search: plans, searches, reads and revises across many turns, rather than answering from a single retrieval pass.
Structured reports: output arrives as a sectioned document with headings and inline sources, suited to complex research questions, market analysis and broad literature scans.
Very long context: the underlying Kimi models carry context windows around a million tokens, so large document sets stay in view across the run.
Open weights: Moonshot publishes the underlying models, so the same capability can be self-hosted where data cannot leave the building.
Limitations & things to watch
It searches the open web, not a peer-reviewed index, so preprints, press releases and secondary coverage carry the same apparent weight as published work unless you check.
Coverage of Chinese-language sources is a strength, but it also skews which material surfaces on some topics.
A structured report reads as more authoritative than the evidence under it, the citation list has to be opened rather than trusted.
Data handling on the hosted consumer service should be reviewed before unpublished or confidential research goes into it.