Recently, Kimi launched the Kimi-Researcher (Deep Research) Agent model based on end-to-end autonomous reinforcement learning technology and opened beta testing.
I applied for beta testing access right away. After an in-depth experience, I found that whether it's everyday life decisions or professional research, Kimi can lay it all out clearly for you.
With just one sentence, it will automatically read materials, break down complex problems, proactively propose research paths, decompose tasks, track information, and finally output a 10,000-word report with rich content and precise sources, and also organize it into a visualized page with clear key points that is easy to read.
Today I'll share with you what exactly Kimi's Deep Search can do? And how Kimi-Researcher achieves SOTA.
01. Hands-on experience
Friends with beta testing access can see the Deep Research button in the dialog box. We input the requirement and select Deep Research here to use this feature.
I chose 7 requirements in different directions to test its deep research feature.
case 1: Assisting complex purchasing decisions
The Xiaomi YU7 launch event is coming soon. I wanted to learn some knowledge about this model in advance, so I told Kimi:
Prompt: Xiaomi YU7.
Kimi will guide me to put forward more specific requirements，and listed a few directions for me to choose from: the Xiaomi YU7's launch price, specific configurations, technical highlights, market competition, and so on.
These points were indeed directions I wanted to learn about, so we could directly have it do comprehensive deep research:
At this point, it will autonomously search and sort out information related to the YU7. The whole process took about twenty to thirty minutes, during which you can completely leave it alone.
In the end, it will organize the information obtained into a detailed deep research report and a visualized report.
Let's first look at the deep research report together. The whole piece is 9,600+ words, with over a hundred citations. The content is comprehensive and detailed, and clear information sources can be seen for all key data.
It fully displayed the several key points mentioned earlier, and indicated the sources in the content. The information sources selected are also relatively authoritative websites in the professional field, such as Autohome, Xiaomi Auto, and Chejiahao.
Compared with the text-based report, the visualized report is more intuitive, and you can see more data comparisons, making it convenient for us to quickly understand the key points.
Open the link to view the completed report: https://www.kimi.com/share/d1eiapsjc3f16p1sumsg
case 2: Planning/optimizing travel itineraries
Prompt: I want to travel from Beijing to Switzerland. Help me plan a one-week itinerary with a budget of 50,000, and try to include all the popular attractions.
Similarly, it generated for me a one-week in-depth Switzerland tour itinerary plan and budget guide, as well as a visualized report.
In the in-depth tour itinerary plan and budget guide, transportation, attractions, accommodation, and food are all introduced in great detail.
In the visualized report, all kinds of information are clear at a glance.
If you don't want to read too much text, you can understand all the content directly by looking at the visualized report.
Open the link to view the complete report:https://www.kimi.com/share/d1eie8mmcu0nio8n7t40
case 3: Financial investment analysis/market research
Prompt: The current status and trends of Chinese games going overseas.
Deep research report:
Visualized report:
Open the link to view the complete report: https://www.kimi.com/share/d1eiapsjc3f16p1sumsg
case 4: Scientific research and literature review
Prompt: Search for literature related to whether tea or coffee has a better refreshing effect.
Besides some official media outlets, I noticed that its content cited quite a lot of Zhihu column content.
If, during academic-related research, the database could be integrated with Kimi's academic search database, that would be even better.
In-depth research report:
Visualization report:
Open the link to view the full report:https://www.kimi.com/share/d1eid6jof8jlmi2au4lg
case 5: Policy analysis and interpretation
Prompt: How much can homebuyers save after the housing provident fund interest rate is lowered?
In-depth research report:
Visualization report:
Open the link to view the full report:https://www.kimi.com/share/d1eid0kjc3f16p1t9sjg
case 6: In-depth information retrieval and analysis
Prompt: Analyze the market competitiveness of the company Moonshot AI.
In-depth research report:
Visualization report:
Open the link to view the full report:https://www.kimi.com/share/d1eidcn37oq3d4ubqqi0
case 7: Professional-grade research report generation
Prompt: Provide an overview of the latest progress in the field of AI large models.
In-depth research report:
Visualization report:
Open the link to view the full report:https://www.kimi.com/share/d1eidh8osk9hkuhggi20
This visualization report is a major highlight of Kimi-Researcher. Various charts and data are presented very clearly, so you can tell at a glance what it is actually saying.
For all of this, you do not need to open dozens of tabs or switch back and forth between materials; you just state your needs, and it can run through the entire research chain.
It would be even better if downloading could be provided~
02. Advantages of Kimi-Researcher
Through the several cases above, we can clearly recognize that Kimi-Researcher's information-gathering ability is really quite good. And its ability to further process the collected information is also among the very best.
It has achieved very good results in tests such as Humanity's Last Exam (HLE) and the xbench benchmark released by Sequoia China. It is currently in the first tier of this type of model.
Its core capabilities can be divided into three levels:
- Active search (WebAgent capability)
It actively uses a browser to search for related content, and then judges for itself which content is useful.
- Multi-document summarization (ultra-long context)
A single research task may involve more than a dozen or even hundreds of materials, with tens of thousands of words of content. Kimi's million-token context capability makes summarizing multiple articles effortless.
- Problem decomposition (multi-step reasoning)
For complex problems, it actively breaks them down into multiple sub-problems, searches, analyzes, and summarizes each separately, and then synthesizes a conclusion. The results produced this way are more detailed.
Traditional Agents require people to write out the process in advance; whereas Kimi-Researcher is entirely based on Kimi's self-developed K-series models and trained through end-to-end Agentic RL. At every step in solving user needs, it continuously adjusts itself based on real-time feedback from the environment.
03. Some sharing
As a self-media person who has long been writing, researching, and producing content, my evaluation of Kimi-Researcher is:
"This is the Agent mode I truly want."
AI is no longer merely a chat partner, but has truly begun to 'work for you' like a human.
Over the past half year, everyone has been discussing Agents, but most remain at the proof-of-concept (PoC) or development tool stage.
The release of Kimi-Researcher marks: the official launch of a To C-layer Agent product. It has achieved a complete closed loop of "goal → path → execution → output".
This is not a "demo", but a productivity tool that can be used immediately and generate value.
It is foreseeable that more and more fields (financial analysis, legal review, medical research) will begin to see similar "Researcher-type Agents".
Original link:Hands-on test of Kimi-Researcher, this is the Agent I truly want!