Answering Common Kimi API Questions and Exploring AI App Development

Why Use a Large‑Model API?

Automation scenarios such as batch processing of large text volumes (translation, rewriting, outline extraction) and integrating model inference into a customer‑service pipeline for classification and automatic replies require programmatic access, which the web assistant cannot provide.

Available Kimi API Endpoints

List available models

Model inference (Chat Completion)

File operations (upload, list, delete, retrieve)

Utility functions (e.g., token‑count calculation)

Account management (balance query, etc.)

Model Variants and Context Windows

Kimi offers three model specifications: moonshot-v1-8k, moonshot-v1-32k, and moonshot-v1-128k. All share the same inference capability; the numeric suffix indicates the maximum context window size in tokens (8 k, 32 k, 128 k), which limits the total length of input plus output per request.

A token is a fragment after tokenisation; for example, the Chinese sentence “我知道了” becomes three tokens: “我”, “知道”, “了”. Token counts do not map directly to word or character counts.

Pricing

Kimi charges only for the Chat Completion endpoint, billing by the total number of tokens (input + output). The pricing table is shown below.

Unlike OpenAI, which differentiates input and output token prices, Kimi applies a unified price, placing its cost at a mid‑range level among Chinese large‑model APIs.

Differences Between API and Web Assistant

The web assistant includes built‑in web‑search capability; the API does not. Developers must call external search services and feed results into the model if needed.

The web assistant uses a proprietary system prompt that influences output. With the API, developers supply their own system prompt, optionally mirroring the web version.

System prompt reference: https://github.com/cssmagic/Awesome-AI/issues/2

Non‑Determinism and Predictability

Provide stricter prompts with detailed background, precise format requirements, and concrete examples.

Set the temperature parameter to a low value (e.g., 0.2) for more focused, deterministic output; higher values (e.g., 0.7) increase randomness.

Cache identical request results to improve response speed and reduce costs.

Obtaining JSON‑Formatted Output

Craft prompts that explicitly request JSON output and include a sample structure; this works in most cases but may require error handling or retries for occasional format deviations.

An undocumented Kimi feature that forces JSON output has been reported (see https://zhuanlan.zhihu.com/p/687898495); it is private and may change.

Providing examples is crucial even when using OpenAI’s JSON mode.

Planned Features

JSON mode : slated for release soon.

Function Calling : core capability for AI agents, prioritized for imminent launch.

Multimodal : image recognition expected within the year.

2‑million‑token context : anticipated first on the web version as a premium feature; API support would likely increase inference costs dramatically, so developers should continue improving Retrieval‑Augmented Generation techniques.

References

Kimi API beta group

Kimi official documentation: platform.moonshot.cn/docs

OpenAI official documentation: platform.openai.com/docs