早鸟价(12 月 25 日截止),订阅期间一直锁定:Go $10/月(原价 $20),Plus 用 Opus 5.5 只要 $30/月(原价 $50)立即订阅 →
文本对话1M 上下文

Kimi K3

通过与 OpenAI 兼容的 Chat Completions API,在线试用 Kimi K3,适用于长上下文代码库、研究资料集合、文档审阅和智能体记忆。

价格 / 100 万 token省 50%
官方AIReiter
输入$3.00$1.50
输出$15.00$7.50
缓存读$0.30$0.15
说明文档
输入
限制回复长度。推理模型也会把思考计入其中。
kimi-k3
输出

A codebase you don't know is a risk, not a hurdle. The plan should burn down that risk in order: understand, run, trace, spike, then make the change small enough that "production-ready" is a property you can demonstrate, not a hope.

Here are the first five steps, with the evidence you collect at each.

Step 1 — Get it building and the baseline green (half day)
Read the README, the manifest(s), the CI config, and any architecture docs. Boot the app locally, run the full test suite, run the linter/typechecker.
Evidence: The exact commands that reproduce a clean build and a passing test suite from a fresh checkout (recorded output, not memory of it). A one-page map you wrote yourself: where the entry points are, the layers, the build/test toolchain, and the main risks to your feature (e.g. "this touches cron jobs and payments").

Step 2 — Trace the smallest end-to-end slice that resembles your feature (half day)
Don't start with your feature. Pick a neighboring one that already works and follow it from entry to persistence. Note the conventions along the way: how errors surface, how config is injected, how logging is done, how tests are written.
Evidence: An annotated trace — file → function → what it does — for that slice. A written list of "conventions I must follow" (not intuitions; things you observed in real code). A named shortlist of the 3–5 files you will actually touch. If you can't produce this trace, you're not ready to write code.

Step 3 — Spike the core path (day 2)
Build the ugly version: hardcoded values, no error handling, no tests. The purpose is to confirm the path you traced in step 2 is real and to surface what you didn't know you didn't know.
Evidence: A working spike demonstrating the feature's central data path, alongside a list of every assumption the spike broke and what you corrected. That correction list is the most valuable document in this whole plan.

Step 4 — Write the contract before the code (half of day 2 / day 3)
Once the spike proves the path, pin down what production needs: the inputs/outputs, the error cases, where it sits in the conventions from step 2. Then write the tests — they'll be red, but they're the specification.
Evidence: A one-to-two-page design doc, an agreed interface/API shape (with the team if there is one), and a red test suite that encodes intended behavior. If you can't write the contract without consulting the code, you haven't finished step 2.

Step 5 — Implement in small, verified increments (days 3–4)
Replace the spike with the real thing in small commits, each one keeping the suite green, leaning on the existing patterns. Run lint/tests/typecheck per commit — CI, not just locally — and exercise the actual path against a real instance (staging or a local environment that isn't stubbed).
Evidence: A branch with progressive commits, each green in CI; coverage on the new code; something that proves it works against reality (a test result, a log trace, a screenshot); and a review by at least one person who knows the codebase. The review counts as evidence — an unfamiliar codebase has tribal knowledge you cannot extract from the files alone.

Steps 6+ would be the things that actually make it "shipped": a migration plan and its rollback, feature flagging, observability, the release and post-release verification. But the first five get you to a reviewed, green, working slice in staging — which is the point at which you can say "this will work in production" with evidence behind it, instead of a guess.

生成耗时42.7s
输入 Token134
输出 Token2,354
token / 秒55.13
TTFT-

运行会以此提示词打开 AI Chat,需要登录并有积分。

模型详情

在 Playground、API 请求和内部工作流中使用相同的模型 key。

模型 ID
kimi-k3
供应商
Moonshot
协议
OpenAI Chat Completions · Anthropic Messages
上下文窗口
1,048,576 Token
最大输出
131,072 Token
输入 Token
150 credits / 1M Token
输出 Token
750 credits / 1M Token
缓存读取
15 credits / 1M Token
缓存写入
-

使用 Kimi K3 你可以做什么

当上下文是瓶颈,而一次请求需要承载大量代码、文档、证据或智能体历史时,选择 Kimi K3。

长上下文审阅

将大型代码库、文档集合或研究证据保留在同一个工作上下文中。

仓库分析

借助更多项目状态可用,梳理文件之间的关系并讨论变更。

研究整合

在生成结构化结论之前,对比来自多份笔记和来源的主张。

智能体记忆评估

审查长工具轨迹和先前决策,找出自动化工作流出错的地方。

Kimi K3 使用场景

最适合那些在提示中保留更多证据可以避免过早分块、检索或丢失项目状态的工作流程。

代码库审阅

在单次请求中分析更多仓库上下文。

长文档集合

审阅合同、政策、报告或研究资料集合。

智能体轨迹分析

检查冗长的工具历史和保留状态。

上下文密集型原型

在构建检索之前,测试更多上下文是否能提升结果。

用 3 步试用 Kimi K3

  1. 01

    选择你的设置

    设置模型支持的响应控制和上传选项。

  2. 02

    发送提示词

    描述任务,添加相关上下文,并查看流式响应和 token 使用情况。

  3. 03

    连接 API

    使用文档中的端点和你的 API key,将同一模型接入你的产品。

使用 Kimi K3 API 构建

管理 API 密钥 →

从交互式测试到生产集成,使用可预测的控制和用量报告顺利过渡。

01熟悉的协议

使用为此模型配置的 API 协议,包括可用的流式传输。

02用量可见性

在每次响应后跟踪输入 token、输出 token 和已消耗的积分。

03模型专属控制

传入支持的生成参数,而不是依赖通用默认值。

04一个账户和余额

通过同一个 AIReiter 账户和计费系统测试并使用受支持的文本模型。

常见疑问

01Kimi K3 最适合做什么?

当一次请求需要大量的代码、文档、研究证据或智能体历史时使用它。

02Kimi K3 可用的上下文窗口有多大?

AIReiter 将 Kimi K3 标注为拥有 1,048,576 token 的上下文窗口;在发送超大请求前,请先验证客户端限制和超时设置。

03我可以使用 OpenAI 风格的客户端调用 Kimi K3 吗?

可以。AIReiter 通过与 OpenAI 兼容的 Chat Completions 端点提供它。

04Kimi K3 如何定价?

当前的输入、缓存读取和输出 token 费率由 AIReiter 显示;请在正式使用前确认。

05什么时候应该改用更小的模型?

对于不需要 Kimi K3 长上下文能力的简短、无状态请求,请使用更轻量的模型。