Last updated: July 2026
Key Takeaways
- Kimi K3 is free to use through the Kimi app and website, but the current privacy policy states that user content is processed for purposes "including training and optimizing our models."
- Unlike ChatGPT, Claude, and DeepSeek, Kimi documents no in-product opt-out for model training as of July 2026, and its policy never names the country where your data is stored.
- Several clauses drawing criticism are industry-standard. The two findings above are not — and the only structural exit from any provider's terms is the open-weights release coming July 27, hardware permitting.
Kimi K3 took the #1 spot on Arena.ai's Frontend Code Arena this week, and nearly every writeup ends the same way: it is free, go try it at kimi.com. That advice is not wrong. It is just incomplete, because none of those writeups link the documents that define what "free" means.
We read them. This is what Moonshot AI's own privacy policy and API terms say you exchange for access, how that compares clause-by-clause with DeepSeek, ChatGPT, and Claude, and what actually protects you — including the one option no closed model can offer.
The Free Tier Everyone Is Recommending
The short version of the news: Kimi K3, Moonshot AI's new 2.8-trillion-parameter flagship, debuted at the top of a major coding leaderboard ahead of the closed models from Anthropic and OpenAI, and the full weights are scheduled for public release by July 27. We covered the ranking, the architecture, and the hardware reality in our Kimi K3 hardware reality check. The result is a wave of first-time visitors to kimi.com, most of them arriving on the strength of a leaderboard screenshot and the word "free."
Free consumer AI is never free in the accounting sense. The question is what the exchange actually is, and the answer is written down. Moonshot publishes a consumer privacy policy for the Kimi app and website, and a separate set of terms for its paid API platform. Both are short reads. Both are specific.
What Kimi's Privacy Policy Actually Says
The consumer document that governs kimi.com and the Kimi apps is the Kimi Privacy Policy, last updated July 7, 2025. The services are provided and controlled by Moonshot AI PTE. LTD., a Singapore-incorporated entity. The load-bearing sentence sits in the section on user content, which covers prompts, audio, images, videos, files, and anything else you input or generate: that content is processed to provide and improve the services, "including training and optimizing our models." The stated legal basis is legitimate interest or consent, depending on your jurisdiction — which means whether you were ever asked depends on where you live.
The collection scope is broader than prompts. The policy lists automatic collection of IP address, device identifiers, conversation and session IDs, and — a detail worth pausing on — clipboard data, where applicable and permitted by your device settings. It also permits training on publicly available information gathered from the open web, with pseudonymization applied where feasible.
Some of the document reads better than critics suggest, and accuracy requires saying so. Voiceprints, which the policy classifies as sensitive personal information, require explicit consent that can be withdrawn. The policy states plainly that Moonshot does not sell personal information. It commits to breach notification, names a dedicated data-protection function, and grants access, correction, deletion, and objection rights. Account deletion is available, permanent, and irreversible, with carve-outs for legal obligations.
The sharing section is where "does not sell" earns its footnotes. Personal information may flow to service providers, to corporate affiliates under common ownership or control, to acquirers in a business transaction, and to public authorities on lawful request under applicable legal standards, including international ones. None of that is unusual. What is unusual comes next.
First, storage. The policy states your information may be transferred to and stored on servers outside your country of residence — and never names a country. Moonshot was founded in Beijing; the consumer service is controlled from Singapore; the storage jurisdiction is simply unstated. The confident claims circulating online about exactly where Kimi data lives are not supported by the document, in either direction. Unstated does not mean China. Unstated means unstated, and for a data-sovereignty decision, an unnamed jurisdiction is itself the finding.
Second, the absence. There is no product-level training opt-out described anywhere in the policy — no toggle, no setting, no checkbox. Independent reviews of the document in early 2026 reached the same conclusion after checking the app itself. The rights the policy grants are exercised through account settings where available or by emailing the company, and which rights you actually hold depends on your jurisdiction. Every major Western competitor, and DeepSeek, now documents a switch. Kimi documents a mailbox.
The policy's most honest sentence is its own advice: in bold text, in the security section, it tells users to take special care deciding what information they send, because absolute security cannot be guaranteed. On that point, we and Moonshot agree completely.
Paying Does Not Change It: The API Terms
A reasonable assumption is that paid access buys stricter handling. For Kimi's API platform, the current Terms of Service for Kimi OpenPlatform say otherwise. Content — defined as both your inputs and the model's outputs — may be used to provide, maintain, develop, support, and improve the services. A customer who wants restrictions on the use of their content for training or improving Moonshot's models "may contact Moonshot AI to discuss available enterprise arrangements."
Read that sentence the way a procurement lawyer would. Training restrictions are not the default, and not a setting. They are a negotiation, available at enterprise scale. That is the inverse of the posture at OpenAI and Anthropic, where API traffic is excluded from training by default and the consumer tier is where the training happens. On Kimi's platform, the meter running does not stop the collection.
One practical caution follows from this: API aggregators and routing dashboards summarize provider policies, but summaries are not contracts. The provider's own documents govern what happens to your tokens — and today, every K3 request ultimately terminates at Moonshot's endpoint regardless of how it is routed. Who else gets to serve those weights, and under what retention terms, is exactly what changes on July 27. That jurisdiction question deserves its own guide, and it is getting one.
How Kimi Compares to DeepSeek, ChatGPT, and Claude
Fairness requires the wide shot. Training on consumer conversations by default is now the industry norm, not a Kimi eccentricity, and every provider in this table discloses data to authorities on lawful request. The differences that matter live in the controls.
| Provider | Trains on consumer chats by default | In-product opt-out documented | Paid API trained by default |
|---|---|---|---|
| Kimi (Moonshot AI) | Yes, per its privacy policy | No — none documented as of July 2026; rights requests go through email | Terms permit content use; restrictions require enterprise arrangements |
| DeepSeek | Yes, per its privacy policy | Yes — an opt-out right in the policy plus a settings toggle; availability varies by region | Not clearly stated; open-platform data handled under separate terms |
| ChatGPT (OpenAI) | Yes, per its data-use policy | Yes — Data Controls toggle and a privacy portal; Temporary Chats excluded | No — API and business tiers excluded unless the customer opts in |
| Claude (Anthropic) | Yes, under the terms in effect since late 2025, unless the user declines | Yes — a Privacy Settings toggle, presented at signup; Incognito chats excluded | No — API and commercial products excluded |
Documented mechanics as of mid-July 2026, per each provider's own published policies, linked in this article. Policies change; verify against the primary documents before relying on any row.
The details behind two of those rows matter. OpenAI's official data-use policy pairs the consumer toggle with a privacy portal request labeled "do not train on my content," and states that business and API customers are excluded from training unless they explicitly opt in. Anthropic's consumer terms update made training on Free, Pro, and Max chats conditional on a user choice presented at signup and changeable in settings — with a significant retention asymmetry attached: five years of retention for users who allow training, roughly 30 days for those who decline. That asymmetry drew criticism when it shipped, and it belongs in any honest comparison. It is also, still, a documented switch — which is more than the Kimi column can say.
DeepSeek's privacy policy grants an opt-out right for model training, and its terms reference a settings switch — "Improve the model for everyone" — whose exact wording and availability vary by region. DeepSeek's policy also does something Kimi's does not: it names where the data lives. You may not like the answer, but it is on the page, and a named jurisdiction can be evaluated. An unnamed one cannot.
What Actually Protects You
Ranked by how much it actually accomplishes, for Kimi or any consumer chatbot.
Do not paste anything you would not want retained. This is the whole game, and Kimi's own policy says as much in bold. Financial documents, health details, client code, anything under NDA — the consumer tier of any provider in the table above is the wrong place for it. We watched this lesson land in April when a lawsuit alleged Perplexity conversations were reaching Meta and Google through embedded trackers, allegedly even in its incognito mode. Terms describe intent; architecture decides outcomes.
Know what does not help. A VPN hides your IP address from your ISP and changes nothing about what a chatbot provider may do with the prompts you type into your account. Incognito browser modes only stop local history. Neither touches the data terms you agreed to, so neither is a fix here, and we will not pretend otherwise.
Use the controls that exist, where you are. If you use ChatGPT, Claude, or DeepSeek, the toggles in the table take two minutes. For Kimi, the documented path is a jurisdiction-dependent rights request by email, and account deletion — permanent — remains the backstop.
Then there is the structural answer, and it is the reason this site covers AI at all. A model running on hardware you control has no privacy policy, because your prompts never leave your network. Consumer-class open models are genuinely capable now — our guide to the best local AI models by VRAM tier maps what runs on everything from an 8GB laptop up, and pairing a local stack with network-level controls like Pi-hole DNS monitoring keeps the whole arrangement observable. And on July 27, K3 itself joins the open side of the ledger: once the weights are public, anyone with sufficient hardware — a real constraint, as our reality check details — can run the #1 frontend model under no terms but their own. That option is what open weights buy, and no toggle at any hosted provider can match it.
Frequently Asked Questions
Is Kimi K3 free to use?
Yes. K3 is available through the Kimi website and apps at no charge, with usage limits, alongside paid subscription tiers and API access priced at $3 per million input tokens and $15 per million output tokens at launch (July 2026). The exchange on the free and consumer tiers is governed by the privacy policy described above, including the model-training clause.
Does Kimi train on your conversations?
The privacy policy states that user content — prompts, files, images, audio, and generated output — is processed to provide and improve the services, explicitly including model training and optimization. The stated legal basis varies by jurisdiction between legitimate interest and consent.
Can you opt out of Kimi using your data for training?
No in-product control is documented as of July 2026. The policy grants jurisdiction-dependent rights — access, deletion, objection, consent withdrawal — exercised through settings where available or by emailing Moonshot. On the API side, the terms state that customers requiring training restrictions may contact the company about enterprise arrangements.
Where does Kimi store your data?
The policy does not say. It states data may be transferred to and stored on servers outside your country of residence, without naming any jurisdiction. The consumer service's controller of record is Moonshot AI PTE. LTD. in Singapore; the company was founded in Beijing. Claims more specific than that are not supported by the published documents.
Is Kimi K3 safe to use?
For general questions, brainstorming, and non-sensitive work, the practical risk profile resembles other consumer chatbots, and the policy includes standard security commitments — encryption, access controls, breach notification. For sensitive personal, financial, medical, or proprietary content, the policy's own bolded advice to be careful what you send is the correct standard, and it applies to every provider in our comparison table, not only this one.
What is the private way to run K3-class AI?
Locally, on hardware you own. Smaller open models already run well on ordinary machines, and a dedicated always-on box is affordable — our mini PC guide for local AI covers hardware tiers and secure network setup. K3's own weights arrive July 27; running the full model at home is out of reach for consumer hardware today, but the distilled and quantized descendants of open flagships are exactly how frontier capability historically reaches machines people actually own.
