Kimi K3 Burns Through $14 of Tokens in an Hour and Still Earns a Grudging Thumbs-Up
by 雪隐_上班了 from juejin.cn/user/143341…
Sharing and aggregation are welcome, but full-text reprinting is not necessary. Please respect copyright; the circle is only so big. If you need it urgently, you can contact me for authorization.
Let's start with the conclusion: This thing is absolutely another "Shoryuken" from China's large model scene, hitting me so hard I'm clutching my wallet with one hand while giving a thumbs-up with the other.
🌟 The Good: Dense with satisfaction, like an intern on cheat codes
1. Gets it right in one shot, almost never blows up
I throw a task at it, and it doesn't rush to write code. Instead, it first builds a mini prototype for verification—like a chef tasting the soup before cooking. Once it confirms the approach is correct, it rolls up its sleeves and gets to work, and finally writes its own test cases to prove it didn't half-ass the job.
The code it delivers is basically copy → run → done. For us old hands, this saves the blood-pressure-spiking time of fixing bugs; but for those newbies who are afraid of mistyping a semicolon, it's practically the "Bodhisattva of Programming"—no more turning the entire codebase into the Leaning Tower of Pisa just because you added a feature.
2. Concurrent multi-file work, like activating shadow clones
When faced with a project needing four or five files, it directly summons 4-5 sub-Agents to work simultaneously, crackling away like a five-stack in an internet cafe. For large projects, it's like hiring a temporary little team for you, and development speed goes straight to lightspeed.
3. Open source is the trump card
After open-sourcing, the community wizards will definitely "buff it up." It's already formidable now, and in the future, it might just become sentient—I'm even looking forward to someone adding a "Boss-Repellent Mode."
4. Performance rivals top-tier closed-source models like Fable 5, and it's "Made in China"
Frontend UI is written with fluid grace, and backend Agent orchestration sits firmly in the world's first tier. The cheekiest part is—while the US keeps trying to choke us off, this thing is essentially handing an "AI Equality Ticket" to every country on the planet that lacks the ability to develop its own large models. That kind of vision, I can only tip my hat.
💸 The Bad: The pleasure is real, but so is the pain
1. Expensive enough to hurt (but compared to foreign options, it's actually a "conscience price")
I bought the 99-yuan package, thinking it would last a week, right? I had it write a moderately complex project, and it burned through 5 hours' worth of tokens in a single hour—then I was forced into a 5-hour "cooldown," like a game skill on cooldown.
Over two afternoons, a week's quota was gone ahead of schedule.
Friendly tip: If your needs are heavy, go straight for the 699 package, or you'll end up like me, staring at your token balance every day with more anxiety than watching the stock market.
2. Using a sledgehammer to crack a nut; even simple tasks get a full doctoral dissertation
I just wanted to write a "Hello World"-level script, and it spent half a minute thinking, listed a 12-step plan, and finally threw in performance optimization suggestions as a bonus…
Dude, I just wanted to fry an egg, and you brought out the kitchen setup for a Manchu-Han Imperial Feast. Tokens just get burned up by this "sense of ceremony."
3. Runs on the first try ≠ flawless
Don't think you can just be a hands-off boss. After single-step debugging, I found that the code it wrote always deviates from the "complete form" I had in mind—like a button being 2px off, or logical details not following my hidden "unspoken rules."
In the end, I still had to pull each one aside for a "heart-to-heart" before it would change. So, a black-box free lunch? Doesn't exist. You are still that worrywart product manager + tester + client.
4. Speed is a bit "zen"
When generating fast, it's like lightning, but on complex tasks, it switches to "slow live-stream" mode.
I suggest brewing a cup of tea, or doing a set of push-ups, and by the time you come back, it'll just be "handing in its paper"—patience is a required course for K3 users.
🧠 Personal Summary
Kimi K3 isn't some "brainless magic lamp," but rather a smart but slightly stubborn, fast-working but money-burning, top-tier capable but overthinking-prone AI partner.
It massively lowers the barrier to programming, allowing non-CS majors to build decent stuff; at the same time, it leaves plenty of "tuning" room for the veterans, proving that we humans haven't been completely replaced yet.
It's not perfect, but it's dazzling enough.
As a member of China's large model army, it deserves this line:
"Awesome, but next time, don't burn through my tokens so fast, okay?" (Wallet is crying, will top up again next time.)