Kimi Code + K2.7 Code Hands-On: Can It Replace Claude Code?

Last week, the two biggest news stories were: Fable 5 (currently the strongest model) was taken offline, and SpaceX went public. Interestingly, both events involve the domestic large model Kimi. First, almost at the same time Fable 5 was taken offline, Kimi released its latest flagship model K2.7 Code[1], which possesses the strongest coding capabilities in this series. The coincidence in timing inevitably creates the impression that Kimi is filling the gap left by its predecessor. Second, SpaceX's listing means that Cursor, the coding agent it acquired for $60 billion, will also go public. One reason for Cursor's high valuation is that it has its own base model Composer, but we now know that the current Composer 2 was not independently developed but is based on Kimi K2.5. Interestingly, Kimi's own coding agent, Kimi Code[2], has also been undergoing intensive updates recently. It claims to be a rival to Claude Code, positioning itself as the best companion for K2.7 Code. These two things together made me very curious about Kimi's new model and coding agent. Although Kimi has never been a household name, it has always enjoyed a good reputation both domestically and internationally, with a considerable user base. The general consensus is that it is capable, affordable, and offers outstanding value for money. I wanted to know: can it really challenge Anthropic and replace Claude Code? So, I immediately got my hands on the Kimi Code + K2.7 Code combo to see how they actually perform.

Coding agents are the most convenient way to use large models on the command line. Claude Code is the pioneer in this field and has the highest market share. Now, almost all large model companies have launched their own coding agents, and Kimi Code is one of them. My biggest impression after using it is that it really looks and feels very similar to Claude Code—they have an extremely high degree of similarity. The first image above shows the startup interface of Kimi Code, and the second shows Claude Code. They look quite alike, don't they? Their operation commands are also compatible. Almost all commands for Claude Code can be run on Kimi Code; I haven't encountered any unsupported ones yet. Even the 'permission approval' prompts during runtime are very similar, appearing with roughly the same frequency. The first image above shows a permission approval request from Kimi Code, and the second from Claude Code. Honestly, if you didn't know, just looking at the UI, you might not be able to tell them apart.

A few days ago, when Fable 5 was released (before it was taken offline), a Japanese netizen used it to generate an ink wash effect[3], where ink naturally spreads, flows, and blends on paper. I found the demonstration very impressive. Since a key feature of Kimi Code is its ability to understand images and videos and generate code based on visual input, my first test was to have it produce this ink wash effect. I submitted this video to Kimi Code and asked K2.7 to generate a similar effect based on the video. I dragged the video into the dialog box and entered the prompt: 'Refer to this video and generate an effect of ink naturally spreading, flowing, and blending on paper.' Kimi Code then broke down the video into a series of screenshots, analyzed them one by one, and finally generated the result shown below[4]. As you can see, it completely replicated the UI from the video and created the ink flow effect. However, because the animation effect was copied from the video rather than generated by a pure algorithm, it is not as smooth as the original. If the video contained a static webpage, I believe it could replicate it exactly. I have posted the generated result online; you can check it here[5]. Kimi Code's visual capability has great practical value. Just think: you can send screenshots and screen recordings directly to the model, saving you the trouble of lengthy and imprecise language descriptions, making it easier to get the desired result.

Kimi Code recently also supports the /goal command. I think this is a very important feature, so I had to test it. The /goal command is used to set a specific goal for the large model. Once the goal is achieved, the model stops running immediately. The reason for introducing this command is that prompts are often too vague, just a general description, causing the model to overthink, initiate multiple rounds of cycles, and constantly ask for user confirmation, which is very time-consuming and costly. With the /goal command, the model has a clear goal and can theoretically run autonomously without user intervention, stopping automatically once the preset condition is met. My test was simple. I found a large image (shown below) with a resolution of 6629 × 5504 pixels and a file size of 2.1MB. Then, I entered the following prompt: '/goal Resize this image so that its file size is just under 120KB.' Note that the prompt starts with the /goal command, followed by a very specific goal. At this point, Kimi Code will ask you to confirm whether you allow it to run completely autonomously. Selecting the first option, 'Switch to Auto and start,' grants full authorization to Kimi Code. There are no operations requiring your approval; you can walk away and do other things. Below is the result. It resized the image to 948 × 787 pixels, resulting in a file size of 119.59KB, which just meets the 'less than 120KB' requirement. The entire process required no human intervention and was very smooth. If you need the large model to run a task for a long time, I highly recommend using the /goal command.

Long-time readers know that I have created an AI test case repository[6], collecting some programming problems, mainly web programming, to test various large models. Running different models through these problems makes it easy to compare them. Below are the test results for K2.7. I have already published the generated results online; you can check them via the links below.

(1) Web UI Refactoring Test[7].

(2) Web 3D Animation Test 'Celestial Body Simulator'[8].

(3) Web Game Test 'Angry Birds'[9].

(4) Web 2D Animation Test 'Maxwell's Demon Experiment'[10].

(5) SVG Image Test 'Pelican Riding a Bicycle'[11].

From these tests, the web capabilities of Kimi K2.7 are solid, and all results are normal. I just feel that the UI could be a bit more refined.

Most of Kimi Code's features are the same as Claude Code's, but it also has some exclusive features. Among them, my favorite is the Kimi Datasource plugin[12]. Kimi Datasource is the official data plugin for Kimi Code. It supports querying financial market data, macroeconomic indicators, company registration records, academic literature, and Chinese laws and regulations using natural language—without the need to manually call APIs or register for data accounts. It is a data source plugin that connects to databases like Tonghuashun and Tianyancha, allowing you to query real-time stock prices, financial reports, academic papers, etc., for free. It's very practical. Installation is simple: enter /plugins, select Marketplace to enter Kimi's plugin marketplace, and choose kimi-datasource. Then, enter /reload to activate the plugin. When using it, enter /kimi-datasource [your question] to query papers, financial reports, and company information. Here are a few examples:

(1) /kimi-datasource Latest research on the side effects of weight loss drugs.

(2) /kimi-datasource The latest income statement of Kweichow Moutai.

(3) /kimi-datasource Who are the shareholders of ByteDance?

This plugin can bring in the latest data for large models and supports multiple data sources. I recommend using it.

After these tests, I have a very good impression of Kimi Code. It is a very useful coding agent and can indeed replace Claude Code. As for the Kimi K2.7 Code model it is paired with, being a trillion-parameter model, it has strong coding capabilities, runs very smoothly, and successfully completed all tasks. What impressed me the most is that its Token consumption is relatively economical. After running all the tests above, only a small portion of my membership quota was consumed. According to the official introduction[13], this is because K2.7's reasoning capabilities have been 'slimmed down,' with Token consumption reduced by an average of 30% compared to the previous version. Kimi's strength has always been its high cost-performance ratio, and this time it has reinforced that advantage, making it worth watching.