Kimi K3 是开放权重，而非开源。Moonshot AI 在 Hugging Face 上以 moonshotai/Kimi-K3 发布了该 checkpoint，采用 Kimi K3 License，这是一份自定义许可证，没有 SPDX 标识符，也不在 Open Source Initiative 批准列表中。Hugging Face 模型卡将 license 字段设置为 other，名称为 kimi-k3。如果你想通过 API 调用该模型，而不是自己运行权重，我们以 moonshotai/kimi-k3 提供该模型。
本文介绍该许可证授予和要求的权利、该 checkpoint 包含的内容，以及如何在 OpenRouter 上调用该模型。
摘要
- Kimi K3 是开放权重。权重是公开的，但 Kimi K3 License 是 Moonshot AI 自己的文本，并未获得 OSI 批准。
- 该许可证默认授予使用、修改、分发和销售的权利。它在规模化使用时增加了两个条件：Model as a Service 收入门槛，以及用户界面署名要求。
- 该 checkpoint 是一个 2.8 万亿参数的混合专家模型，每个 token 激活 1040 亿参数，以 MXFP4 存储。
- 在 OpenRouter 上，模型 ID 是
moonshotai/kimi-k3。它接受文本、图像和视频输入，拥有 1,048,576 token 的上下文窗口，并支持推理强度、工具调用和结构化输出。没有:free变体。
开放权重与开源
这两个术语回答的是不同的问题。
开放权重意味着发布方公开了训练好的参数，你可以下载并查看它们。附加在这些参数上的许可证可以是发布方自行撰写的任何文本。
开源，按照 Open Source Initiative 对该术语的用法，意味着许可证符合 Open Source Definition 并出现在 OSI approved list 上。像 Apache-2.0 和 MIT 这样的许可证拥有 SPDX 标识符，并且在该列表上。
Kimi K3 符合第一个定义，但不符合第二个。
Kimi K3 License 授予了什么
许可证文本将本软件定义为模型权重、参数、配置文件、推理与训练代码以及相关文档。它免费授予使用、复制、修改、合并、发布、分发、再许可和销售本软件副本的权利，以及运行、部署、微调本软件并基于其创建衍生作品的权利。
第 1 节设定了两项基本条件。你必须在软件的所有副本或实质性部分中包含版权声明和许可声明，并且你的使用必须遵守适用的法律法规。
Kimi K3 License 在大规模使用时的要求
在特定阈值之上还适用另外两项条件。
模型即服务门槛。第 2 条将模型即服务定义为：向第三方提供对语言模型推理或微调的访问权限，例如通过 API，且其方式使第三方能够对输入、参数或训练数据行使实质性控制。该定义不包括将模型能力嵌入特定功能或框架之中的终端用户产品，也不包括将请求转发至他人托管的模型。如果你或你的任何关联方经营模型即服务业务，且你的合计收入在任何连续 12 个月内超过 2000 万美元，那么在为任何商业目的使用本软件或其衍生作品之前，你必须与 Moonshot AI 签订单独协议。
用户界面署名。第 3 条适用于以下情形：本软件或衍生作品被用于月活跃用户超过 1 亿或月收入超过 2000 万美元的商业产品或服务。在此情况下，该产品必须在其用户界面上显著展示“Kimi K3”。
豁免。第 4 条将两种情形豁免于上述两项条件之外。第一种是内部使用，定义为不向第三方提供本软件、其输出或其底层能力的用途。第二种是通过 Moonshot AI 的官方产品或认证推理合作伙伴进行访问。
第 5 条是保证免责声明。本软件及其输出按现状提供，不附带任何形式的保证。
在基于这些权重发布产品之前，请先阅读许可证。本文描述的是截至 2026-09-11 发布的文本，不构成法律建议。
检查点包含的内容
以下数据来自 Hugging Face 模型卡。
| 项目 | 数值 |
|---|---|
| 总参数量 | 2.8T |
| 每 token 激活参数量 | 104B |
| 专家数 | 共 896 个，每 token 激活 16 个 |
| 注意力机制 | Kimi Delta Attention 与 Attention Residuals |
| 视觉编码器 | MoonViT-V2，401M 参数 |
| 上下文长度 | 1,048,576 tokens |
| 权重格式 | MXFP4 权重，MXFP8 激活，量化感知训练 |
| 许可证 | Kimi K3 许可证 |
| Hugging Face 仓库 | moonshotai/Kimi-K3 |
模型卡将 Kimi K3 描述为一个原生多模态模型，能够理解文本、图像和视频。模型卡列出思考始终启用，并带有一个 reasoning_effort 字段，该字段接受 low、high 和 max，默认为 max。对于多轮对话和工具调用，模型卡指出，API 返回的完整助手消息，包括推理内容和工具调用，必须回传至 messages。
模型卡列出了三个已发布配方的本地服务栈：vLLM、SGLang 和 TokenSpeed。我们尚未对这些服务栈进行基准测试，本文也不涉及自托管。2.8 万亿参数的 checkpoint 需要多节点服务硬件。如果你没有这样的硬件，托管 API 就是运行该模型的方式。
如何在 OpenRouter 上调用 Kimi K3
模型 ID 为 moonshotai/kimi-k3。2026-09-11 的目录记录列出了以下属性。
| 属性 | 值 |
|---|---|
| 输入模态 | 文本、图像、视频 |
| 输出模态 | 文本 |
| 上下文长度 | 1,048,576 tokens |
| 推理 | 默认启用，efforts low、high、max，默认 max |
| 支持的参数 | tools、tool_choice、response_format、structured_outputs、reasoning、reasoning_effort、include_reasoning、max_tokens、temperature、top_p、top_k、seed、stop、logprobs、top_logprobs、logit_bias、frequency_penalty、presence_penalty、repetition_penalty、min_p |
| 变体 | moonshotai/kimi-k3:batch 用于 Batch API。没有 :free 变体。 |
定价
我们不为该模型设定统一价格。每个提供商端点都会列出自己的提示词、补全和缓存读取价格，模型页面会显示当前的价格表。2026-09-11 时，目录中有 19 个端点用于 moonshotai/kimi-k3。提示词价格从每百万 token $1.80 到 $6.00 不等，补全价格从每百万 token $9.01 到 $22.50 不等，缓存读取价格从每百万 token $0.21 到 $0.60 不等。Moonshot AI 自己的端点列出的是 $3.00 提示词、$15.00 补全和 $0.30 缓存读取。最大输出长度也因端点而异，从 16,384 token 到 943,718 token 不等。
提供商端点所支持的参数也各不相同。2026-09-11 时，有两个端点未列出 tools，有两个未列出 structured_outputs。如果你的请求依赖某个参数，请将 provider.require_parameters 设为 true，这样我们就只会路由到支持该请求中所有参数的端点。完整的路由控制项请参见 提供商选择。
推理
Kimi K3 默认返回推理内容。通过我们的 API，你可以使用 reasoning 对象来控制它。reasoning.effort 对该模型接受 low、high 和 max。将 reasoning.exclude 设为 true 会从响应中移除推理文本，而模型仍会进行推理。将 reasoning.enabled 设为 false 会要求端点跳过推理。
我们于 2026-09-11 针对 moonshotai/kimi-k3 将每种配置各运行了一次，使用了一个简短的算术提示词。默认请求、low 和 max 请求各自返回了一个 reasoning 字段，以及一条带有答案的 reasoning_details 条目。exclude 请求返回了答案，但没有推理字段。enabled: false 请求没有返回推理字段，而是返回了一个更长的答案，在 content 字段中展示了其推理过程。模型卡称思考始终处于启用状态，而单次调用无法表明该端点究竟是停止了内部推理，还是仅仅停止了返回推理内容。请将 enabled: false 视为对该端点的一项请求，而非一种保证。
对于多轮对话和工具调用，请将完整的 assistant 消息传回 messages，其中包括 reasoning_details 和 tool_calls。我们的推理指南涵盖了 跨轮次保留推理。
带图片的示例请求
该请求发送一张图片和一个问题，要求 high 推理强度，并打印答案。它使用 TypeScript SDK。
import { OpenRouter } from '@openrouter/sdk';
const openRouter = new OpenRouter({
apiKey: process.env.OPENROUTER_API_KEY ?? '',
});
const result = await openRouter.chat.send({
chatRequest: {
model: 'moonshotai/kimi-k3',
messages: [
{
role: 'user',
content: [
{
type: 'image_url',
imageUrl: { url: 'https://example.com/screenshot.png' },
},
{
type: 'text',
text: 'Summarize this screenshot and list the open questions.',
},
],
},
],
reasoning: {
effort: 'high',
},
stream: false,
},
});
if (result instanceof ReadableStream) {
throw new Error('Expected a non-streaming response');
}
console.log(result.choices[0].message.content);我们于 2026-09-11 使用一张公开照片运行了一个等效请求。该端点返回了对该图片的一句话描述、一个 reasoning 字段，以及一个 usage 对象，其中包含 96 个 prompt tokens、119 个 completion tokens 和 87 个 reasoning tokens。图片输入在 图像理解中有所介绍，视频输入则在 视频输入中有所介绍。
工具调用与结构化输出
模型记录列出了 tools、tool_choice、response_format 和 structured_outputs。工具定义遵循 工具调用中的形式。JSON Schema 响应遵循 结构化输出。由于对这些参数的支持因端点而异，当请求依赖这些参数时，请将 provider.require_parameters 设置为 true。
提示词缓存
2026-09-11 时，每个 moonshotai/kimi-k3 端点都列出了缓存读取价格。当请求命中提供商的提示词缓存时，缓存的提示词 token 将按该端点的缓存读取费率计费，而非其提示词费率，并且响应会在 usage.prompt_tokens_details.cached_tokens 中报告这些 token。关于各提供商的缓存机制，请参见 提示词缓存。
Kimi 早期版本
Kimi K2 和 Kimi K2.5 采用修改版 MIT 许可证。两者都在 MIT 基础上增加了一条条款，要求月活跃用户超过 1 亿或月收入超过 2000 万美元的商业产品，必须在用户界面上显著展示模型名称。Kimi K3 许可证保留了该署名条款，并新增了模型即服务收入门槛。
我们的目录还列出了 moonshotai/kimi-k2.6、moonshotai/kimi-k2.7-code、moonshotai/kimi-k2.5、moonshotai/kimi-k2-thinking 和 moonshotai/kimi-k2。本文仅介绍 Kimi K3。
常见问题
Kimi K3 是开源的吗？
不是。Kimi K3 是开放权重的。Moonshot AI 在 Hugging Face 上以 Kimi K3 License 发布了权重，这是一份自定义许可证，不在 Open Source Initiative 批准列表之中。该许可证授予使用、修改、分发和销售该模型的广泛权利，但附带两项在营收和用户数阈值上生效的条件。
Kimi K3 License 允许什么？
该许可证授予使用、复制、修改、合并、发布、分发、再许可和销售该模型副本的权利，以及运行、部署、微调和创建衍生作品的权利。你必须在该软件副本中保留版权和许可声明，并遵守适用法律。
Kimi K3 License 增加了哪些条件？
两项。如果你或你的关联方经营 Model as a Service 业务，且在任何连续 12 个月内的合计营收超过 2000 万美元，则你在商业使用前需要与 Moonshot AI 另行签订协议。如果基于该模型构建的商业产品月活跃用户超过 1 亿，或月营收超过 2000 万美元，其用户界面必须显著展示 Kimi K3。内部使用以及通过 Moonshot AI 官方产品或认证推理合作伙伴进行的访问，均不受这两项条件约束。
我可以下载 Kimi K3 的权重吗？
是的。该 checkpoint 已发布在 Hugging Face 上的 moonshotai/Kimi-K3。这是一个 2.8 万亿参数的混合专家模型，以 MXFP4 格式存储，每个 token 激活 1040 亿参数。
如何在 OpenRouter 上调用 Kimi K3？
向 https://openrouter.ai/api/v1/chat/completions 发送一个 chat completion 请求，模型 ID 为 moonshotai/kimi-k3。该端点接受文本、图像和视频输入，返回文本，并支持 reasoning、tools 以及结构化输出参数。
OpenRouter 上有 Kimi K3 的免费版本吗？
没有。目录中列出了 moonshotai/kimi-k3 以及 Batch API 的单独 moonshotai/kimi-k3:batch 条目。两者都不是免费的。定价由各提供商设定，因此请查看 模型页面以了解当前费率。
Kimi K3 is open-weight, not open-source. Moonshot AI published the checkpoint on Hugging Face as moonshotai/Kimi-K3 under the Kimi K3 License, a custom license that has no SPDX identifier and is not on the Open Source Initiative approved list. The Hugging Face model card sets the license field to other with the name kimi-k3. If you want to call the model through an API instead of running the weights yourself, we serve it as moonshotai/kimi-k3.
This post covers what the license grants and requires, what the checkpoint contains, and how to call the model on OpenRouter.
Summary
- Kimi K3 is open-weight. The weights are public, but the Kimi K3 License is Moonshot AI’s own text and is not OSI-approved.
- The license grants use, modification, distribution, and sale by default. It adds two conditions at scale, a Model as a Service revenue gate and a user-interface attribution requirement.
- The checkpoint is a 2.8 trillion parameter mixture-of-experts model with 104 billion active parameters per token, stored in MXFP4.
- On OpenRouter the model ID is
moonshotai/kimi-k3. It accepts text, image, and video input, has a 1,048,576-token context window, and supports reasoning effort, tool calling, and structured outputs. There is no:freevariant.
Open-weight versus open-source
The two terms answer different questions.
Open-weight means the publisher released the trained parameters so you can download and inspect them. The license attached to those parameters can be any text the publisher writes.
Open-source, as the Open Source Initiative uses the term, means the license meets the Open Source Definition and appears on the OSI approved list. Licenses like Apache-2.0 and MIT have SPDX identifiers and are on that list.
Kimi K3 meets the first definition and not the second.
What the Kimi K3 License grants
The license text defines the software as the model weights, parameters, configuration files, inference and training code, and associated documentation. It grants, free of charge, the rights to use, copy, modify, merge, publish, distribute, sublicense, and sell copies of the software, and to run, deploy, fine-tune, and create derivative works from it.
Section 1 sets two baseline conditions. You must include the copyright notice and the permission notice in all copies or substantial portions of the software, and your use must comply with applicable laws and regulations.
What the Kimi K3 License requires at scale
Two further conditions apply above specific thresholds.
Model as a Service gate. Section 2 defines Model as a Service as giving a third party access to language model inference or fine-tuning, for example through an API, in a way that lets the third party exercise meaningful control over the inputs, parameters, or training data. The definition excludes end-user products where model capabilities are embedded within specific features or harnesses, and it excludes relaying requests to models hosted by others. If you or any of your affiliates operate a Model as a Service business, and your aggregate revenue exceeds 20 million US dollars over any consecutive 12 months, you must enter into a separate agreement with Moonshot AI before using the software or its derivatives for any commercial purpose.
User-interface attribution. Section 3 applies when the software or a derivative is used in a commercial product or service with more than 100 million monthly active users or more than 20 million US dollars in monthly revenue. In that case the product must prominently display “Kimi K3” on its user interface.
Exemptions. Section 4 exempts two cases from both conditions. The first is internal use, defined as use that does not make the software, its outputs, or its underlying capabilities available to third parties. The second is access through Moonshot AI’s official products or certified inference partners.
Section 5 is the warranty disclaimer. The software and its outputs are provided as is, without warranty of any kind.
Read the license before you ship a product built on the weights. This post describes the text as published on 2026-09-11 and is not legal advice.
What the checkpoint contains
The figures below are from the Hugging Face model card.
| Fact | Value |
|---|---|
| Total parameters | 2.8T |
| Activated parameters per token | 104B |
| Experts | 896 total, 16 activated per token |
| Attention | Kimi Delta Attention and Attention Residuals |
| Vision encoder | MoonViT-V2, 401M parameters |
| Context length | 1,048,576 tokens |
| Weight format | MXFP4 weights, MXFP8 activations, quantization-aware trained |
| License | Kimi K3 License |
| Hugging Face repository | moonshotai/Kimi-K3 |
The model card describes Kimi K3 as a native multimodal model that understands text, images, and video. It lists thinking as always enabled, with a reasoning_effort field that accepts low, high, and max, defaulting to max. For multi-turn conversations and tool calls, the card states that the complete assistant message returned by the API, including reasoning content and tool calls, must be passed back in messages.
The card lists three local serving stacks with published recipes, vLLM, SGLang, and TokenSpeed. We have not benchmarked these stacks and this post does not cover self-hosting. A 2.8 trillion parameter checkpoint requires multi-node serving hardware. If you do not have that hardware, a hosted API is the way to run the model.
How to call Kimi K3 on OpenRouter
The model ID is moonshotai/kimi-k3. The catalog record on 2026-09-11 listed these properties.
| Property | Value |
|---|---|
| Input modalities | text, image, video |
| Output modality | text |
| Context length | 1,048,576 tokens |
| Reasoning | enabled by default, efforts low, high, max, default max |
| Supported parameters | tools, tool_choice, response_format, structured_outputs, reasoning, reasoning_effort, include_reasoning, max_tokens, temperature, top_p, top_k, seed, stop, logprobs, top_logprobs, logit_bias, frequency_penalty, presence_penalty, repetition_penalty, min_p |
| Variants | moonshotai/kimi-k3:batch for the Batch API. No :free variant. |
Pricing
We do not set one price for the model. Each provider endpoint lists its own prompt, completion, and cache-read price, and the model page shows the current table. On 2026-09-11 the catalog had 19 endpoints for moonshotai/kimi-k3. Prompt prices ranged from $1.80 to $6.00 per million tokens, completion prices from $9.01 to $22.50 per million tokens, and cache-read prices from $0.21 to $0.60 per million tokens. Moonshot AI’s own endpoint listed $3.00 prompt, $15.00 completion, and $0.30 cache read. Maximum output length also varied by endpoint, from 16,384 tokens to 943,718 tokens.
Provider endpoints also differ in which parameters they support. On 2026-09-11, two endpoints did not list tools and two did not list structured_outputs. If your request depends on a parameter, set provider.require_parameters to true so we only route to endpoints that support every parameter in the request. See provider selection for the full set of routing controls.
Reasoning
Kimi K3 returns reasoning by default. Through our API you control it with the reasoning object. reasoning.effort accepts low, high, and max for this model. reasoning.exclude set to true removes the reasoning text from the response while the model still reasons. reasoning.enabled set to false asks the endpoint to skip reasoning.
We ran each configuration once against moonshotai/kimi-k3 on 2026-09-11 with a short arithmetic prompt. The default, low, and max requests each returned a reasoning field and one reasoning_details entry with the answer. The exclude request returned the answer with no reasoning fields. The enabled: false request returned no reasoning fields and a longer answer that showed its working in the content field. The model card says thinking is always enabled, and a single call cannot show whether the endpoint stopped internal reasoning or only stopped returning it. Treat enabled: false as a request to the endpoint, not a guarantee.
For multi-turn conversations and tool calls, pass the complete assistant message back in messages, including reasoning_details and tool_calls. Our reasoning guide covers preserving reasoning across turns.
Example request with an image
This request sends an image and a question, asks for high reasoning effort, and prints the answer. It uses the TypeScript SDK.
import { OpenRouter } from '@openrouter/sdk';
const openRouter = new OpenRouter({
apiKey: process.env.OPENROUTER_API_KEY ?? '',
});
const result = await openRouter.chat.send({
chatRequest: {
model: 'moonshotai/kimi-k3',
messages: [
{
role: 'user',
content: [
{
type: 'image_url',
imageUrl: { url: 'https://example.com/screenshot.png' },
},
{
type: 'text',
text: 'Summarize this screenshot and list the open questions.',
},
],
},
],
reasoning: {
effort: 'high',
},
stream: false,
},
});
if (result instanceof ReadableStream) {
throw new Error('Expected a non-streaming response');
}
console.log(result.choices[0].message.content);We ran an equivalent request with a public photograph on 2026-09-11. The endpoint returned a one-sentence description of the image, a reasoning field, and a usage object with 96 prompt tokens, 119 completion tokens, and 87 reasoning tokens. Image input is covered in image understanding and video input in video input.
Tool calling and structured outputs
The model record lists tools, tool_choice, response_format, and structured_outputs. Tool definitions follow the shape in tool calling. JSON Schema responses follow structured outputs. Because support for these parameters varies by endpoint, set provider.require_parameters to true when the request depends on them.
Prompt caching
Every moonshotai/kimi-k3 endpoint on 2026-09-11 listed a cache-read price. When a request hits the provider’s prompt cache, the cached prompt tokens are billed at that endpoint’s cache-read rate rather than its prompt rate, and the response reports them in usage.prompt_tokens_details.cached_tokens. See prompt caching for how caching works per provider.
Earlier Kimi releases
Kimi K2 and Kimi K2.5 use a modified MIT license. Each adds one clause to MIT that requires commercial products with more than 100 million monthly active users or more than 20 million US dollars in monthly revenue to prominently display the model name on the user interface. The Kimi K3 License keeps that attribution clause and adds the Model as a Service revenue gate.
Our catalog also lists moonshotai/kimi-k2.6, moonshotai/kimi-k2.7-code, moonshotai/kimi-k2.5, moonshotai/kimi-k2-thinking, and moonshotai/kimi-k2. This post covers Kimi K3 only.
FAQ
Is Kimi K3 open source?
No. Kimi K3 is open-weight. Moonshot AI published the weights on Hugging Face under the Kimi K3 License, a custom license that is not on the Open Source Initiative approved list. The license grants broad rights to use, modify, distribute, and sell the model, with two conditions that apply at revenue and user-count thresholds.
What does the Kimi K3 License allow?
The license grants the rights to use, copy, modify, merge, publish, distribute, sublicense, and sell copies of the model, and to run, deploy, fine-tune, and create derivative works. You must keep the copyright and permission notices in copies of the software and comply with applicable law.
What conditions does the Kimi K3 License add?
Two. If you or your affiliates operate a Model as a Service business and your aggregate revenue exceeds 20 million US dollars over any consecutive 12 months, you need a separate agreement with Moonshot AI before commercial use. If a commercial product built on the model has more than 100 million monthly active users or more than 20 million US dollars in monthly revenue, its user interface must prominently display Kimi K3. Internal use and access through Moonshot AI’s official products or certified inference partners are exempt from both.
Can I download the Kimi K3 weights?
Yes. The checkpoint is published at moonshotai/Kimi-K3 on Hugging Face. It is a 2.8 trillion parameter mixture-of-experts model stored in MXFP4, with 104 billion parameters active per token.
How do I call Kimi K3 on OpenRouter?
Send a chat completion request to https://openrouter.ai/api/v1/chat/completions with the model ID moonshotai/kimi-k3. The endpoint accepts text, image, and video input, returns text, and supports the reasoning, tools, and structured output parameters.
Is there a free variant of Kimi K3 on OpenRouter?
No. The catalog lists moonshotai/kimi-k3 and a separate moonshotai/kimi-k3:batch entry for the Batch API. Neither is free. Pricing is set per provider, so check the model page for the current rates.