Kimi: Do you trust a cheap Chinese model with your data?

Every AI newsletter you follow spent last week telling you Kimi K3 is the cheap frontier model that finally caught the US labs.

Running it as an individual, they're right about the model. But they're skipping the only question a business owner should ask first.

Moonshot's K3 landed July 16 at 2.8 trillion parameters, the biggest open-weight model anyone's released for free, and it trades punches with GPT-5.5 and Claude Opus 4.8 on coding. CNBC says it "rivals OpenAI, Anthropic."

Kimi won’t be a real frontier model in my mind until it hacks an infrastructure provider.

The benchmark nobody's running is what happens to your customer list once it's sitting on a server in Beijing.

Kimi is built by Moonshot AI. A Chinese company. Which means Chinese law gets a say in what happens to your data, and you don't.

Why almost nobody asks where their prompts actually end up

• They see "open weights" and think "private." Two very different things.
• They paste client contracts into a free hosted app without opening the terms once.
• They treat cheap as free, and miss that the price just moved from your wallet to your harvested data.

But is Kimi actually as good as everyone says?

If your machine can run it then mostly yes, with one asterisk. Everyone I've watched run it splits the same way.

What they rate it for:

• #1 on Arena.ai's Frontend Code Arena
• The long-context crowd is the loudest group in its favor. People feeding it 100-page reports and whole repos say it tracks how files relate better than anything else they've used.

What the same people complain about:

• The hallucination numbers going around put K3 at 51%, up from 39% on K2.6. Accuracy climbed too, 46% from 33%. Paradoxically it gets more right and invents more wrong at once. I'd treat those as directional until someone reruns them properly.
• Give it a vague instruction and it just goes. No question, no check-in.

Good enough that you'll start feeding it the real stuff.

Where your data goes

1. It won't name a country. The policy you agree to at kimi.com will only say your information "may be transferred to and stored on servers located outside of your country of residence."

Their developer terms do name a country, Singapore, twice. Nice, for a company whose people and offices are in Beijing.

You’re not fooling us.

2. Your content trains the model, with no switch to stop it.

• What counts as user content: "prompts, audio, images, videos, files, and any content you input or generate."
• What happens to it: processed "including training and optimizing our models."
• Whether they need your permission: the legal basis may be "our legitimate interests or your consent, depending on your jurisdiction." In some countries they don't have to ask.
• How you switch it off: you don't. ChatGPT, Claude and DeepSeek all have a toggle. Kimi's is an email request, and the paid API is opted in by default unless you negotiate an enterprise deal.

Retention runs "as long as necessary," with no deletion clock in the sentence.

3. The paperwork your compliance person wants doesn't exist.

No SOC 2, no signed HIPAA BAA. If you sell into healthcare or finance, forget Kimi.

The walls between users have already slipped once. In April 2026, Kimi served one person's private resume, name and phone number included, to a completely different user mid-session. Verified, and it triggered legal action (in China).

As of early 2026, Kimi was pulling roughly 3.5 times more unsanctioned enterprise traffic than DeepSeek ever did. Mostly code and financial projections.

We’ve seen this play before

Rewind eighteen months. DeepSeek was the cheap Chinese model everyone rushed to. Sound familiar?

Its privacy policy said user data sat on servers in China.

Once your data is on a Chinese server, Chinese law applies. Article 7 of the National Intelligence Law of 2017 requires organizations to support and cooperate with state intelligence work.

South Korea's regulator then found DeepSeek shipping user data to ByteDance servers, the actual text people typed into prompts included. No consent, no disclosure. Suspended in the country.

Italy, Australia, Taiwan and a stack of US agencies pulled it off their devices inside a few months.

Has anyone caught Kimi doing that? No. There is no public report of K3 smuggling your data anywhere. Yet.

But nobody had caught DeepSeek either, right up until someone opened it up and looked.

Kimi is the same setup: cheap, Chinese, hosted in China, free enough that nobody asks hard questions.

Chinese rules apply either way, and we all know how that government treats data it can reach. Moonshot doesn't get a vote on that.

"We haven't found the problem yet" and "there is no problem" are not the same sentence.

Beijing’s finest ⬆️

I'd keep it off anything client-facing until someone you trust has poked at it.

Decide what's OK to send

Not everything you touch is sensitive. Use that.

• Fine for the hosted app: public marketing copy, blog drafts, generic brainstorming, anything you'd happily post publicly anyway.
• Never for the hosted app: customer lists, financials, unreleased product info, proprietary code, anything under an NDA.
• If you need Kimi's quality on sensitive work: self-host the weights or rent them from a Western host. Or run a US frontier model where the jurisdiction and the contract are both ones you understand.

But if you’re like me and you’re busy running a company, you know you’re gonna slip something in there accidentally, that ends up in Xi Jin Ping’s cousin’s business plan.

So better to just get smart on how you actually use your tokens more efficiently in a SAFE model.

We’ll cover that next week. Tell your friends ❤️

(I meant share this with the only other founder you care about)

PS - do it now.