A new AI model launches, and two days later the maker shuts the door: no new subscriptions, servers glowing. That's exactly what just happened to Moonshot AI — a Chinese start-up whose model Kimi K3 triggered a rush its own infrastructure can't handle.
It's possibly the most honest advertising a model can get. You must have done something right when people pile in faster than you can bolt in graphics cards.
What Kimi K3 actually is
Kimi K3 is Moonshot AI's new flagship, out since July 17. Two numbers place it:
- 2.8 trillion parameters. Parameters are the dials tuned during training — roughly: the more of them, the more knowledge fits in. That makes K3 the largest open model to date (more on that in a moment).
- Frontier level. In independent tests K3 plays right at the top — in one coding discipline it even landed ahead of Claude Fable 5, the current US front-runner. Overall it sits just behind. For a model from China, built despite the US chip bans, that's a statement.
"Open" here means: Moonshot plans to release the full model weights for download (announced for late July). Anyone may grab them and run the model themselves. The opposite of ChatGPT, where you only get in through OpenAI's door. What "open but not free" means in practice we unpacked elsewhere — in short: those 2.8 trillion parameters absolutely will not run on your laptop. That takes server hardware in the six-figure range.
Benchmarks are one thing. Our in-house silliness test is another: asking for an essay on why a PommDöner beats a Lahmacun — yes, we run this on every new model. K3 delivered without hesitation, titled "The Untouchable Crown: Why the PommDöner Puts the Lahmacun in the Shade". A closing taste: you need nothing more — "except perhaps an Ayran and a little humility before this culinary greatness". Confident, punchy, no waffle. Passed.
The rush — and the back door
Demand sextupled within 48 hours, according to Moonshot. The result: new subscriptions halted, existing customers served first, fresh slots returning "in waves."
Sounds locked out. But it's only half true.
Because the halt applies only to Moonshot's own app. Through the programming interface (API — the direct machine access, billed per token instead of a monthly plan) K3 stays reachable, among others via OpenRouter, the one-stop gateway to many models. There you pay per use — currently around 3 US dollars per million input tokens, 15 per million output tokens. A token is roughly a syllable; a million output tokens are therefore about five novels.
So the back door is open — but it sticks a little too. OpenRouter itself currently warns: upstream capacity is limited, and the model may return frequent "429" errors (the polite code for "too many requests, try again in a moment"). No surprise — OpenRouter forwards requests to that very same overloaded provider. The bottleneck follows you right up to the back door. With a bit of luck you still get your answer; reliable it is not, right now. Anyone dead set on trying K3 today gets in through the API, even without one of the scarce subscription slots — but needs some patience. Once the weights are out and other providers host the model, that will ease.
And what's free?
The Kimi app and kimi.com have a free tier — chatting, file uploads, web search, all with daily limits. Plenty for trying it out. The paid plans (currently roughly 19 to just under 200 dollars a month, depending on your appetite) are meant for steady use. Except: it's precisely the new sign-up that's jammed right now. The more reliable way to test-drive it at the moment is the API — costs a few cents, works instantly.
Do I need to switch providers now?
Short answer: no. Relax.
If ChatGPT, Claude or Gemini handle your daily grind, a new record model from China changes nothing about that for now. K3 is impressive, but no reason to upend working habits — especially while it's gasping for air itself.
K3 gets interesting for two groups. First: anyone who codes a lot and wants a strong, much cheaper model — looking east has been worth a shot for a while (see also MiniMax). Second: anyone who wants to run a model themselves for privacy or control reasons — that's exactly what open weights are for, even if "run it yourself" here means data center rather than living room.
For everyone else, Kimi K3 is above all one thing: further proof that the AI race no longer runs in Silicon Valley alone. The striking part isn't that China is catching up. It's how fast — so fast that even the winners can't keep pace, not even at reordering servers.