Last night, I tried Kimi K2, and its front-end capabilities are extremely strong! It can replace Claude Code's default model, driving Claude Code with K2, so there's no need to worry about account bans anymore. With its super cost-effective pricing, I'm directly enjoying it! First, I'll look at the test results, and later I'll attach a tutorial on how to use K2 in Claude Code.

Those who can’t wait for the construction 🚧 can see here: mp.weixin.qq.com/s/S5SKpzQ8GFG...…

Let's start with Master Zang's classic prompts. This time, PPT logic is added to make the entire page support PPT-like page turning. Many friends want PPT prompts, and now they are here. The effect is beyond expectations. Even Gemini and Opus 4 often have problems exceeding the page height. K2 succeeded at once, and also added rebound and easing effects to each card, and added animation to the charts, which is very suitable for displaying report content.

Then let's take a look at Master Zang's daily signature website. Basically, it's just a matter of looking at the simple front-end capabilities and the aesthetic quality of building a web page. Let's look at the results directly. The quality and effect are basically the same as Claude Sonnet 4. All functions can be used, and he also did a simple copywriting layout. I breathed a sigh of relief when I saw this. The front-end capabilities of many domestic models can't even handle this. Now I can finally test the weight.

Then there was a visual ability test project, which asked him to do a dot matrix animation. At that time, except for DeepSeek 0528, all the prompts failed and basically could not run. Looking at the results directly, it ran correctly for the first time, and there was no problem with the animation mode and shape selection switching above. The other adjustment options below did not take effect for the first time, but after telling him to fix it once, it was fixed, and the night mode could also be handled normally.

Let's make a product with complex logic that can really work, such as a minimalist Markdown note editor. The result generated by K2 for the first time was very well handled in terms of functions. Basically all functions are available and effective, without bugs. The rendered Markdown style was a bit touching, so I told him that the style rendering of various formats was too bad, let's change it again. Then it was basically fixed, and all styles were fine. Except for the background color of the table, this is easy to say. It was beyond expectations that such a complex function and style processing was done like this for the first time.

How to use K2 in Claude Code This time, the K2 model is not only compatible with the common Open AI calling method, but also compatible with Anthropic's model calling. This means that you can replace K2 with the model that provides services for Anthropic, and you don't have to worry about account suspension and fees. You can use Claude Code completely.

Of course, there are also ways to do it without such hassle. If you have installed Claude Code before, you can directly download the project (github.com/LLM-Red-Team/kimi-c...…）的 sh file) and open it. It will ask you to enter your API Key, and then you are done.

I often tell my friends recently. Overseas Agent programming products are in full swing, with new ones appearing every day. China still does not have a model that meets the requirements of tool calling and front-end capabilities to drive Agents. I feel locked in and unconsciously a little pessimistic. But now that I have tried this open source K2, I feel that I have regained my confidence. I realize that it fills a key gap: engineering practicality. K2 is certainly not the strongest model in the world, but it is another domestic open source model that makes me feel "can be used for work".

Additional explanation: If errors are reported frequently, it may be due to these two problems: First, try these two commands to change the environment variables: export ANTHROPIC_BASE_URL=api.moonshot.cn/anthropic export ANTHROPIC_API_KEY=$api_key Then you need to top up at least 50 yuan to get 1 TPM to use it

Continue to update. At present, the TPM of K2 has been greatly increased. Friends who had problems with TPM just now can try again. The default output length has been adjusted to 32K

New tests: