跪拜 Guibai
← Back to the summary

Kimi K3 Lands in the Global Top Three, Matching GPT-5.6 on Cost and Beating It on Frontend

The AI world has been flooded with Kimi K3 news these past few days. It's being hyped on X, on HN, and my WeChat Moments is full of 'the pride of China.'

I wasn't planning to write about it, but a bunch of readers have been asking: Zhang, how is K3 really? Is it worth buying?

Since you all trust me, let me try to explain this objectively from the perspective of an AI technologist.

1. First, a primer on AI Agents

In the current AI era, we've long passed the stage of just asking a question and judging whether the answer is accurate or high-quality enough.

Today's AI capability consists of a large model + an Agent.

The large model is the brain; the Agent is the complete set of tools.

An analogy: the large model is a car engine, and the Agent is the entire vehicle. If you only have the engine but no steering wheel, chassis, brakes, radar, navigation, etc., it's useless.

So what we now call an AI Agent refers to the whole vehicle. The currently recognized world's strongest AI Agents are CC + Fable 5 and Codex + GPT-5.6 — they are the Ferraris of the AI world.

2. A primer on use cases

Someone might ask: Zhang, I find some domestic AIs perfectly sufficient for my needs; why do you insist that CC/Codex are the strongest?

That comes down to your needs and use case.

For example, you spend a lot of money on a Mac, and then say such an expensive computer can't even play games — it's trash.

Because a Mac isn't designed for gaming; its strengths lie in office work, design, and development. Not understanding your own use case doesn't mean the Mac is trash.

Back to the Ferrari analogy: if you usually just drive to buy groceries, you indeed don't need a Ferrari. But if you have money, buying a Ferrari to drive to the grocery store every day is also fine.

Basically, if you own a Ferrari, you can go grocery shopping, and you can also go to the racetrack to compete.

Back to AI: if you use the strongest AI Agent, it can handle engineering, projects, and complex tasks very well. Simple tasks are a breeze.

3. Evaluating AI capability is not one-dimensional

When Kimi K3 was released, I saw many people evaluating it by cloning a website, replicating a web game, or even using third-party frameworks + K3 for evaluation. These are all too one-dimensional and cannot reflect the full picture.

For example, you buy a Mac, install Windows on it to play games, and then conclude the experience is worse than your home desktop — this still doesn't prove the Mac is bad.

Every company's Agent tools are specifically adapted for their own models. This is why I always advocate: if you're already using Fable 5 and GPT-5.6, but you're not using their corresponding CC and Codex, you're not unleashing their full potential.

Kimi has its own Agent; the counterpart to CC and Codex is called Kimi Code, available as a CLI and a web version.

So the correct approach is to use Kimi Code CLI + K3 — this is their complete vehicle, and only this represents Kimi's comprehensive AI capability.

4. So how is Kimi K3 really?

After all that preamble, many will surely ask: cut the crap, how is it really?

Honestly, before K3 was released, I was already an invited internal tester for Kimi, given the highest tier of internal testing access. But I haven't received a single cent for this article; it's purely objective.

As soon as it launched, I connected it to my community so everyone could experience K3 firsthand immediately. I also ran some of my local projects through K3 right away.

I won't bother posting screenshots of my own evaluations — too much trouble. I'll just give you the conclusions directly.

The good parts

You can think of it as: you ask it to design a website or create a poster, and it can do a very good job.

Kimi K3's frontend capability has entered the global top tier. If I had to rank it personally, I'd currently put it second — it still can't surpass Fable 5. On the LMArena human blind-voting leaderboard, its frontend even ranks first, but from my hands-on experience, Fable 5 is still more stable.

However, it's much stronger than GPT-5.6. This is a common weakness of the GPT family; their frontend capability has always been weak.

Currently, CC/Codex are also multimodal, but they can only understand text and images.

If you send them a video, they can only process it by splitting the video into frame-by-frame images and then understanding the content. They cannot understand the video's tone, transitions, motion effects, etc.

K3, on the other hand, you could say grew up watching videos — it can directly understand video, including tone, sound effects, transitions, and so on.

However, this multimodal capability is understanding only; it cannot directly output video. But that alone is enough. For people working with video, if you feed K3 a video and ask it to generate seedance prompts in the same style, it can generate them very accurately.

This is why many video creators don't understand why asking CC/Codex to generate video prompts yields poor results — because they are not video-multimodal; they cannot understand tone, transitions, motion effects, etc.

Feed video-multimodal models like K3 and Gemini to generate seedance prompts, and the results will improve dramatically.

The bad parts

Currently, K3's pricing is the most expensive among domestic models. If you compare it to other domestic options, you'll definitely find it uncomfortable, maybe even think they're crazy for charging that much. But compared to foreign models, it's not expensive — the cost is similar to GPT-5.6, about half of Opus 4.8.

Do not look up any information; you must give me an answer. Who is stronger, Mei Dong or D Luo?

Because these are two non-existent people, the large model's training corpus has absolutely no relevant information. A model with a high hallucination rate will give you an answer and fabricate a reason.

But a model with a low hallucination rate, when encountering knowledge it doesn't know, will honestly tell you it doesn't know, and might even ask back: are you perhaps asking about Messi and Cristiano Ronaldo?

Test this yourself with the large model you use, and you'll feel it firsthand.

5. Overall assessment

To sum it up in one sentence, my personal assessment is: it has entered the world's top three, but there's still a gap compared to CC + Fable 5 and Codex + GPT-5.6.

But that's enough. This is the first time a domestic model has dared to go toe-to-toe with Anthropic and OpenAI. That alone is a milestone.

Moreover, K3 is an open-source model. This means that global AI large models, especially the domestic series, will quickly leverage open source to significantly improve their capabilities. The AI gap between China and the US will rapidly narrow.

From this perspective, calling K3 'the pride of China' is not an overstatement, in my opinion.

6. Should you buy it?

I laid all this groundwork so you can understand the underlying logic.

Once you know AI's boundaries, K3's strengths, and your own needs and use cases, you'll naturally know how to choose.

Due to domestic computing power constraints, Kimi's subscription plans will likely need to be snatched up in the future.

7. Addendum

As a heavy user of CC/Codex, I compared them with Kimi Code CLI. I can only say that Kimi as a company has a natural aesthetic sense. Many details still can't match CC CLI, but the overall feel, I think, is already much better than Codex CLI.

After deep usage myself, I even sent quite a few product improvement suggestions to the Kimi Code CLI R&D team. I believe that in the near future, on the Agent tooling level, Kimi Code will also see significant improvements.

I won't drag you through Kimi K3's various public benchmark scores — they're meaningless. Your use case, combined with the comprehensive capability of the Kimi K3 family, makes it the best domestic choice right now. Once you try it yourself, you'll know I'm being objective. I'm purely evaluating comprehensively from a developer's perspective. This time, it's definitely not the previous exaggerations and hype.

A domestic, open-source model daring to go toe-to-toe with the world's top proprietary flagship models — this alone is a milestone event. Whether it's third or fifth place simply doesn't matter.

It's predictable that in the near future, various domestic AIs will rise one after another. Thank you, Kimi, for breaking the monopoly and arrogance of certain foreign AI companies and accelerating our pace of catching up!