Kimi K3: specs, price, the license, and who reads your chats

The Morse teamPublished October 26, 20269 min read

Kimi K3 is Moonshot AI's open-weight flagship model, launched in July 2026: 2.8 trillion parameters in total, 104 billion active on each token, a context window of 1,048,576 tokens, and native image and video understanding. Its weights have been on Hugging Face since July 27, 2026 under the Kimi K3 License, and Moonshot's API charges $3 per million input tokens and $15 per million output tokens.

We read Moonshot's model card, its license, its launch post, its API price list, its membership pricing, and the privacy policies and terms for each Kimi service on October 22, 2026. Benchmark scores are Moonshot's unless we say otherwise. Moonshot deserves credit for releasing weights at this size, and its launch post names the two closed models it still trails.

Kimi K3 specs at a glance

Moonshot calls K3 "the world's first open 3T-class model," though the count it gives is 2.8 trillion.

Is Kimi K3 Chinese?

Yes. Moonshot AI is based in Beijing, and the privacy policy linked from kimi.com names Beijing Moonshot Technology Co., Ltd. as the company behind Kimi (Kimi). Kimi's international service is run by a separate company: its terms name NOVASCENT PRIVATE LIMITED, "a company registered in Singapore" (Kimi), and the API is sold by Moonshot AI PTE. LTD., also in Singapore (Kimi API terms). Which company holds your chats depends on which Kimi you use, as the privacy section below shows.

How good is Kimi K3?

Close to the best closed models on agent and coding work, by Moonshot's count, and behind them overall by its own admission. From the model card, at max effort:

• Terminal-Bench 2.1: 88.3, against 88.0 for Claude Fable 5 and 88.8 for GPT-5.6 Sol.
• BrowseComp: 91.2, the highest in Moonshot's table.
• DeepSWE: 67.5, against 70.0 for Fable 5 and 73.0 for GPT-5.6 Sol.
• HLE-Full with tools: 56.0, against 63.0 for Fable 5.
• SWE-Marathon: 42.0, ahead of every model listed, though Moonshot's own footnote says Fable 5 "hit fallbacks on 35% of the tasks in our evaluation."

Moonshot ran most of its coding tests in its own Kimi Code harness, and other labs' scores come from their own setups, so the rows aren't strictly like for like. Epoch AI, which tests models independently, puts K3 19th of the 253 models on its capabilities index and estimates its training compute at 2.0e25 FLOP (Epoch AI).

How much does Kimi K3 cost?

On Moonshot's API, as of October 22, 2026, Kimi K3 costs $3.00 per million input tokens, $0.30 per million cached input tokens, and $15.00 per million output tokens (Kimi API), and Moonshot says there's "no tiering by context length" (Kimi API). Writing a prompt into the cache costs $3.00 per million tokens for a five-minute cache and $6.00 for a one-hour cache.

K3 is the expensive model on Moonshot's list. Kimi K2.6 costs $0.95 in and $4.00 out with a 262,144-token window, so K3 runs about three times the input price and nearly four times the output price. To call K3 at all, your account needs a top-up of at least $1, and the total you've paid in sets your rate limits (Kimi API). The old address, platform.moonshot.ai, now redirects to platform.kimi.ai.

Is Kimi K3 free?

The weights are free to download. Using K3 through Moonshot mostly isn't. Kimi's membership page lists a $0 plan called Adagio and four paid ones, as of October 22, 2026:

Kimi's help center says K3 draws on each plan's shared monthly credit pool, alongside Deep Research, Kimi Code, and the rest, and that credits don't roll over (Kimi). Only the Max and Ultra cards list "Up to 1M tokens of context for ultra-long conversations." We couldn't find a page from Moonshot that says how much K3 the free Adagio plan includes. Even the older K2.6 isn't free in Kimi's chat: the credit rules say "In Chat, K2.6 also consumes membership credits" (Kimi).

What does the Kimi K3 License allow?

Almost anything, with three conditions. Some of the top Google results for "kimi k3 open weights license" describe it as a plain MIT license. The license is Moonshot's own text. It grants the MIT permissions (use, copy, modify, sell, fine-tune) and then adds:

1. Keep the notice. The copyright and permission notice go with every copy.
2. Big API businesses need a deal. If you run a "Model as a Service" business, meaning you give others access to inference or fine-tuning "with meaningful control over the inputs, parameters, or training data," and you and your affiliates take in more than $20 million over any 12 months, you must "enter into a separate agreement with Moonshot AI" before any commercial use.
3. Big products show the name. A product with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display "Kimi K3" prominently in its interface.

Conditions 2 and 3 don't apply to internal use, or to K3 "accessed through Moonshot AI's official products or certified inference partners." An app that only embeds the model in its own features, or only relays requests to a model someone else hosts, isn't a Model as a Service under the license.

The second condition arrived with K3. The license for Kimi K2.6 is a "Modified MIT License" with only the display rule (Hugging Face). For a person or a small company, K3's terms work much like MIT. For a large hosting company, they don't. This isn't legal advice.

Can you run Kimi K3 locally?

Only on server hardware. The repository holds 96 weight files that add up to about 1.56 TB, by our sum of the file sizes Hugging Face lists, before any memory for a long context. Moonshot recommends vLLM, SGLang, or TokenSpeed to serve it and publishes no minimum hardware. The MXFP4 format keeps the files far smaller than 2.8 trillion parameters would be at 16 bits, but it's still a job for a rack of data-center GPUs, not a desktop.

What happens to your chats on Kimi?

Kimi keeps them and, by default, may train on them. Each Kimi service has its own policy, and they don't say quite the same thing. We couldn't find a Moonshot page that says which countries each consumer policy covers, so if you use Kimi, the policy linked when you sign up is the one you've agreed to.

The international policy says Kimi processes "prompts, audio, images, videos, files, and any content you input or generate" to improve its services, "including training and optimizing our artificial intelligence ('AI') models." Its terms add that opting out "does not require us to delete or remove Content already used in prior training cycles." The policy gives no retention period for chats, only "as long as necessary," and Kimi's help center says that after you delete your data, "certain network logs (such as operation timestamps and account identifiers)" are kept "for a specified period" under the Cybersecurity Law (Kimi). The same policy says Kimi's sharing with ad networks "qualifies as the sale of personal information" under some US state laws. It says the data for that advertising is "primarily collected" through cookies and similar technologies.

The API pages contradict each other. Kimi's help center says data sent through the API "is not used to train or improve Kimi's models" (Kimi). The API's terms of service, last updated July 30, 2026, say Moonshot "may use Content to provide, maintain, develop, support, and improve the Services," that customers who want limits on training should "contact Moonshot AI to discuss available enterprise arrangements," and that "unless otherwise expressly agreed in writing, Customer Content may be used for the foregoing purposes" (Kimi API terms). The terms are the contract you accept, so read them as the rule.

None of this applies to the weights. Once K3 runs on someone else's servers, Moonshot's policies don't reach the chat, and whoever runs the model decides what's kept. We cover the same split for another Chinese lab's models in is DeepSeek safe.

How to use Kimi without Moonshot keeping your chats

On kimi.ai or kimi.com, your chats sit with NOVASCENT, a Singapore company, or with Beijing Moonshot in China. Training stops only after you email in, and chats already used stay in past training runs. On Morse, you pick Kimi without a Kimi account of your own, your history stays in your browser, and Morse has zero data retention agreements with its inference, search, and other providers, so they keep nothing. Check the model menu for the Kimi versions in the catalog today.

• Private, the default, so the host keeps nothing. Open-weight models like Kimi, DeepSeek, and GLM run here, under those zero data retention agreements. When a provider confirms zero retention on a reply, Morse verifies it and rejects any reply that lacks it. Elsewhere it relies on the provider's word. Read what zero data retention covers.
• E2EE, when you'd rather trust a check than a policy. Moonshot's help center and its API terms disagree about training, and E2EE rests on a check your browser runs instead. Your browser verifies the enclave's hardware attestation, then seals the message so only that enclave can open it, and the reply is decrypted only on your device, while Morse's servers relay ciphertext they can't read. Kimi is one of the open-weight models that run this way, in enclaves that third-party providers operate. The guarantee holds as long as the enclave hardware and attestation work as designed. E2EE runs only on-device tools, so search the web from a Private chat. E2EE comes with Pro, $19 a month, and our guide walks through each enclave check.
• Anonymous, to test Moonshot's own comparison. Moonshot says K3 trails the top closed models. To check for yourself, ask GPT, Claude, or Gemini the same question in Anonymous mode. Morse's proxy strips your account, email, and IP address, so the provider gets the question without knowing it's you. It may keep the question for a limited time for abuse checks, under its own policy.

No Morse chat waits on an opt-out request. Morse doesn't train AI on what you write, sell your data, or show ads (privacy policy). Your history is encrypted in your browser and locked when you sign out, and Morse can't restore it, so if you clear the browser the chats are gone. Keep a copy of anything you'll need. To see how other labs handle training, read which AI companies train on your data.

Start free to browse the model menu, or choose Pro for end-to-end encrypted chats with open-weight models.