What Is the Kimi K2 Family?Kimi K2 家族是什么？
This page covers the Kimi K2 family, not one version number. The original K2, released July 11, 2025, is Moonshot AI's open-weight agentic model: a 1-trillion-total-parameter Mixture-of-Experts with 32B activated parameters per token, trained on 15.5 trillion tokens, with the MuonClip optimizer reported as achieving zero loss spikes in training. Its stated job is long-horizon agentic coding, tool use and multi-agent orchestration — not chat. Do not confuse it with Kimi the consumer app: Kimi is the chat product, K2 is the model family underneath that class of work.本页讲的是 Kimi K2 家族，不是某一个版本号。初版 K2 于 2025-07-11 发布，是月之暗面的开源权重 agentic 模型：1T 总参数的 MoE（混合专家）、每 Token 32B 激活参数、15.5T tokens 训练量，MuonClip 优化器口径为训练零 loss spike。它的定位是长程 agentic 编码、工具调用与多智能体编排——不是聊天。不要把它和消费端产品 Kimi 混淆：Kimi 是对话 App，K2 是做这类工作的模型家族。
The family has since evolved: a K2 Thinking (thinking variant), K2.6 and K2.7-Code and others have followed — but version dates are disputed across roundups and this page does not pick one quietly. In particular, K2.6's release date is given as both April 20, 2026 and October 3, 2026 in different roundups; treat any single K2.6 date you see elsewhere as one source's claim, not settled fact. Weights are released in the Modified MIT licence direction, and context is about 256K tokens. Pricing is deliberately not reduced to one API number here: K2 is served by Moonshot AI and by third-party API providers at different rates — check the official site and your provider.家族此后一直在演进：K2 Thinking（思考版）、K2.6、K2.7-Code 等先后出现——但各版本日期在横评中说法不一，本页不悄悄采信其一。尤其 K2.6 的发布日期在不同横评中有 2026-04-20 与 2026-10-03 两种说法；别处看到的任何单一 K2.6 日期都只是一家之言、不是定论。权重为 Modified MIT 许可方向，上下文约 256K tokens。价格本页有意不写成单一 API 价：K2 由月之暗面与第三方 API 供应商以不同价格提供——请查官网与你的供应商。
K2's honest niche is agentic engineering work: long coding sessions where the model must call tools, run commands, recover from errors and coordinate sub-agents, with open weights you can self-host. It is not a general chat upgrade and not a reasoning-showpiece purchase: if your job is a single-turn answer, the consumer Kimi app or a frontier chat model is the shorter path. Judge the family by agentic coding benchmarks and by total serving cost, not by headline parameter counts.
K2 实在的定位是 agentic 工程活儿：长程编码会话——模型要调工具、跑命令、从错误中恢复、协调子智能体，且权重开源可自托管。它不是通用聊天升级、也不是推理炫技之选：如果你的活儿是单轮问答，消费端 Kimi App 或前沿聊天模型是更短的路。评判这个家族要看 agentic 编码基准与总服务成本，而不是 headline 参数数。
— 中国吧 AI Tools Hub Editorial Team
Who Should Use It?谁适合用？
✓ Good Fit For适合以下场景
- Long-horizon agentic coding with tool use需要调工具的长程 agentic 编码
- Multi-agent orchestration experiments多智能体编排实验
- Self-hosted open-weight deployments开源权重自托管部署
- Teams benchmarking open models on coding agents在编码智能体上评测开源模型的团队
✕ Not Ideal For不适合以下场景
- Single-turn chat and casual Q&A单轮聊天与随手问答
- Anyone wanting one settled price and one settled version想要一个定价、一个定论版本的人
- Workloads needing a much longer context than about 256K tokens需要远超约 256K 上下文的工作量
Getting Started快速开始
1. Decide the family member first: original K2, Thinking, K2.6 or K2.7-Code1. 先定家族成员：初版 K2、思考版、K2.6 还是 K2.7-Code2. Check the licence and serving route: self-host the weights or pick an API provider2. 查许可与服务路线：自托管权重，还是选 API 供应商3. Benchmark on your own agentic coding task, not on headline scores3. 在你自己的 agentic 编码任务上跑基准，别只看 headline 分数4. Price the full session cost — tokens, tool calls and retries4. 按整会话成本计价——token、工具调用与重试都算上Key Features核心功能
- 1T MoE, 32B Active1T 参数 MoE、32B 激活 — Original K2: 1 trillion total parameters as a Mixture-of-Experts, 32B activated per token — sparse capacity without dense cost.初版 K2：1T 总参数 MoE，每 Token 激活 32B——稀疏容量、不付稠密成本。
- 15.5T Tokens Training15.5T tokens 训练 — Trained on 15.5 trillion tokens, with the MuonClip optimizer reported at zero loss spikes during training.15.5T tokens 训练量，MuonClip 优化器口径为训练零 loss spike。
- Built for Agentic Coding为 agentic 编码而建 — Long-horizon coding, tool calling and multi-agent orchestration are the design target — not single-turn chat.长程编码、工具调用与多智能体编排是设计目标——不是单轮聊天。
- Modified MIT Licence DirectionModified MIT 许可方向 — Open weights under a Modified MIT licence direction — check the exact licence text for your deployment.开源权重，Modified MIT 许可方向——部署前查许可原文。
- About 256K Context约 256K 上下文 — Context around 256K tokens — long enough for large codebases in one session, not unlimited.上下文约 256K tokens——一次会话装得下大型代码库，但不是无限。
Pros & Cons优缺点
✓ Pros优点
- Open weights you can self-host and audit开源权重、可自托管、可审计
- Designed for the agentic coding loop: tools, errors, retries, sub-agents为 agentic 编码循环而设计：工具、错误、重试、子智能体
- A real family line (Thinking, K2.6, K2.7-Code) to track, not a one-off一条真实的家族线（思考版、K2.6、K2.7-Code）可追踪、不是一次性发布
- Sparse MoE keeps serving cost below dense-model levels稀疏 MoE 让服务成本低于稠密模型
✕ Cons缺点
- Version dates are disputed — K2.6 alone has two competing dates版本日期有分歧——仅 K2.6 就有两个相互竞争的日期
- No single settled API price across providers各供应商之间没有统一的 API 定价
- Not a chat product — wrong tool for single-turn Q&A不是聊天产品——单轮问答用它就是用错工具
- Context around 256K tokens caps very large single-session jobs约 256K 上下文限制超大单会话任务
Use Cases应用场景
Popular ways people use it.常见用法：
💻 Autonomous Coding Sessions自主编码会话
An agent runs a multi-hour refactor: reading code, running tests, fixing failures, reporting back.智能体跑数小时重构：读代码、跑测试、修失败、回禀结果。
🧩 Multi-Agent Orchestration多智能体编排
A coordinator agent splits a build across sub-agents for frontend, backend and tests.协调智能体把构建拆给前端、后端与测试子智能体。
🏠 Self-Hosted Model Serving自托管模型服务
A team serves K2 weights on its own GPUs under the Modified MIT licence direction.团队在自有 GPU 上按 Modified MIT 许可方向部署 K2 权重。
📊 Open-Model Benchmarking开源模型评测
A lab compares K2 family members against other open agentic models on identical coding tasks.实验室在相同编码任务上对比 K2 家族成员与其他开源 agentic 模型。
Similar AI Tools相似 AI 工具
The consumer side lives on its own page: Kimi is the chat app for end users, while K2 is the model family for builders. If your job is chatting with an AI, go to Kimi; if it is building agents that code, stay with the K2 family.消费端在另一页：Kimi 是给最终用户的对话 App，K2 是给 builder 的模型家族。要和 AI 聊天去 Kimi；要造会写代码的智能体，留在 K2 家族。