Skip to main content

Moonshot AI (Kimi) Provider Guide

Moonshot AI's Kimi API is a Tier-2 catalog provider: OpenAI-wire-compatible, so its integration is one JSON file (src/lib/providers/catalog/moonshot-ai.json) rather than hand-written code. That file is the source of truth for everything on this page. The one behavioural setting it carries is the replayReasoningContent quirk, explained under Key Facts.

Verification status: this entry is docs-verified only, not yet live-verified. The vendor's GET /v1/models needs a key (an unauthenticated GET answers HTTP 401), so the model ids come from the vendor's public docs at https://platform.kimi.ai/docs/models and the roster has not been checked. evidence.liveMatrix is null until someone runs the live capability matrix with a real key (see Verification status below). No account was created and no API key was used to build this entry.

Key Facts​

  • Provider id: moonshot-ai (aliases moonshot, kimi)
  • Protocol: OpenAI-compatible (/chat/completions). The Quickstart says: "Kimi API lets you interact with Kimi models and is compatible with both the OpenAI and Anthropic API formats."
  • Base URL: https://api.moonshot.ai/v1. The docs and console live on platform.kimi.ai and call the product the Kimi API; the API host is api.moonshot.ai.
  • Default model: kimi-k3
  • Models in catalog: 4 — the models the Model List names under "Multi-modal Model". The chat request schema in the vendor's OpenAPI file names the same four ids.
  • Streaming: supported. The streaming guide documents stream responses as Content-Type: text/event-stream (SSE) ending with data: [DONE].
  • Tool calling: true — the tool-calls guide documents the tools parameter and tool_calls (including tool calls in streaming output), and the Kimi K3, Kimi K2.7 Code and Kimi K2.6 pages describe tool calling for their models. The parameter reference says kimi-k3 supports tool_choice values auto / none / required, while kimi-k2.6 and kimi-k2.7-code "return an error if it is passed" for required.
  • Tools while streaming: true, per the tool-calls guide section "Handle tool_calls in Streaming Output"
  • Structured output: supported — response_format accepts json_object and json_schema (response_format guide). For kimi-k2.6 the same guide states: "kimi-k2.6 occasionally behaves unstably with complex schemas; for example, $ref may return a Markdown code block, oneOf may be ignored, and partial=true may output fields outside the schema."
  • Structured output + tools together: not declared (false) — no combined probe was run without credentials
  • Embeddings: not declared
  • Thinking: not declared (false) on this entry. The vendor's Thinking Models page documents reasoning_content, reasoning_effort for kimi-k3 and the thinking parameter for kimi-k2.6; the generic catalog provider does not send reasoning_effort or thinking, so NeuroLink's thinkingLevel is not mapped to them.
  • Reasoning replay (quirks.replayReasoningContent: true): the Thinking Models page says "For K3, this is required in multi-turn conversations and tool-call loops", and that for kimi-k2.7-code "you must therefore (not optionally) keep the reasoning_content of historical assistant messages in messages as-is". The quirk sends the assistant turns' reasoning_content back on later requests. It has not been exercised against the live API.
  • Billing: no-free-tier — see Billing below
  • Key format: not declared

Quick Start​

1. Get an API key​

  1. Visit: https://platform.kimi.ai/console/api-keys and sign in, then create and copy an API key (https://platform.kimi.ai/docs/overview describes signing in to the Kimi API Platform and creating a key under API Keys)
  2. Create the key on platform.kimi.ai: the errors page states that keys issued on platform.kimi.ai are independent from keys issued on other regional Kimi platforms, and that "Mixing keys across platforms returns 401." (https://platform.kimi.ai/docs/api/errors)
  3. Top up before use: "To prevent abuse, you need to recharge at least $1 to start using, and when your cumulative recharge reaches $5, you will receive a $5 voucher." (https://platform.kimi.ai/docs/pricing/limits); the Kimi K3 page states "Kimi K3 is a flagship model: it is unlocked after a successful top-up (minimum $1)." (https://platform.kimi.ai/docs/guide/kimi-k3-quickstart)
  4. Set MOONSHOT_AI_API_KEY in your .env file (MOONSHOT_API_KEY, the variable Kimi's own docs use, is also read)

2. Configure​

export MOONSHOT_AI_API_KEY=your-api-key
export MOONSHOT_AI_MODEL=kimi-k3   # optional — overrides the default model
export MOONSHOT_AI_BASE_URL=https://api.moonshot.ai/v1   # optional — proxy or self-hosted gateway

3. Use it​

import { NeuroLink } from "@juspay/neurolink";

const neurolink = new NeuroLink();

const result = await neurolink.generate({
  input: { text: "Explain context windows in one paragraph." },
  provider: "moonshot-ai",
  model: "kimi-k3",
});

console.log(result.content);
# CLI
npx @juspay/neurolink generate "Hello" --provider moonshot-ai

Per-request credentials work as they do for the other catalog providers:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "moonshot-ai",
  credentials: { moonshotAi: { apiKey: process.env.MOONSHOT_AI_API_KEY } },
});

Billing​

The Recharge and Rate Limiting page says: "To prevent abuse, you need to recharge at least $1 to start using, and when your cumulative recharge reaches $5, you will receive a $5 voucher." The Compare with Other Kimi Products page describes the Kimi API Open Platform as "pay-as-you-go billing with no subscription plan"; the Troubleshooting page repeats "pay-as-you-go with no subscription plan". The entry records no-free-tier.

The Troubleshooting page also answers "Can I try the models before topping up?" with "You can run a minimal test in the Playground to confirm whether a model and prompt fit your scenario".

The limits page lists rate limits by cumulative recharge amount. Tier0 (cumulative recharge $1) is listed as concurrency 1, RPM 3, TPM 500,000 and TPD 1,500,000.

Models​

ModelContextVision$/M in · out (cached)Notes
kimi-k3 ⭐1,048,576yes$3.00 / $15.00 (cached $0.30)NeuroLink default. Model List: "Kimi's most capable model to date". Cache write $3.00 (TTL 5min) or $6.00 (TTL 1h). "K3 always has thinking mode enabled".
kimi-k2.7-code262,144yes$0.95 / $4.00 (cached $0.19)Fallback. Model List: "Kimi's dedicated coding model." "Kimi K2.7 Code does not support non-thinking mode."
kimi-k2.7-code-highspeed262,144yes$1.90 / $8.00 (cached $0.38)Fallback. Model List: "High-Speed version of Kimi K2.7 Code model, with output speed of approximately 180 Tokens/s and up to 260 Tokens/s in short context scenarios". "the same model as Kimi K2.7 Code". Same page: "(Currently, the resource is limited, and the experience of the high-speed model may be slightly fluctuate，we are gradually increasing the resource.)"
kimi-k2.6262,144yes$0.95 / $4.00 (cached $0.16)Fallback. Model List: "Supports both visual and text input, thinking and non-thinking modes, and dialogue and Agent tasks." "thinking is on by default, can be disabled". response_format guide: "kimi-k2.6 occasionally behaves unstably with complex schemas; for example, $ref may return a Markdown code block, oneOf may be ignored, and partial=true may output fields outside the schema."

Context windows and prices come from the pricing page, retrieved 2026-09-29. Its K3 table columns are "Cached Input Price", "Input Price" and "Output Price"; its K2 table columns are "Input Price (Cache Hit)", "Input Price (Cache Miss)" and "Output Price". The catalog records the cache-hit price as cachedInput and the cache-miss price as input. Prices are per 1M tokens and "exclude applicable taxes".

vision: true for the four models comes from the vision guide, which names kimi-k3, kimi-k2.6, kimi-k2.7-code and kimi-k2.7-code-highspeed in its description of the Kimi Vision Model. It also states that for K3 "Vision input does not support public image URLs. Use base64 or ms://<file-id>" (Kimi K3 page).

Output limits stated by the vendor: for kimi-k3, max_completion_tokens "defaults to 131072 and can be set up to 1048576" (Kimi K3 page); the Troubleshooting page gives "the maximum output length is 1024*1024 - prompt_tokens" for kimi-k3 and "the maximum output length is 256*1024 - prompt_tokens" for kimi-k2.7-code and kimi-k2.6. No per-model maxOutputTokens is recorded in the catalog.

models.defaultContextWindow (262,144) and models.defaultMaxOutputTokens (131,072) apply to model ids outside the catalog and are placeholders: 262,144 is the context window the pricing page lists for kimi-k2.7-code, kimi-k2.7-code-highspeed and kimi-k2.6, and 131,072 is the max_completion_tokens default the vendor states for kimi-k3. The vendor publishes no figure for such ids.

The Model List's Deprecated Models section names kimi-k2.5, the moonshot-v1 series, the kimi-k2 series, kimi-latest and kimi-thinking-preview; they are not in the catalog.

Fixed sampling parameters: the parameter reference lists temperature, top_p, n, presence_penalty and frequency_penalty as "Cannot be modified" for kimi-k3, kimi-k2.7-code and kimi-k2.6, and states that "Fixed" means "the parameter cannot be modified: passing any other value returns an error, so do not pass it explicitly."

Fallback order when the default is unavailable: kimi-k2.7-code → kimi-k2.6 → kimi-k2.7-code-highspeed. The order is the catalog's own choice. The runtime fallback model name the loader derives (fallbacks[1]) is kimi-k2.6.

Verification status​

Tier-2 onboarding requires evidence before a provider is accepted, and pnpm run verify:provider-onboarding gates it in CI. This is what the catalog currently records for Moonshot AI — docs-verified only, not live-verified:

ProbeResult
Rosterunauthenticated GET https://api.moonshot.ai/v1/models answered HTTP 401 on 2026-09-29, so the roster was not read; the four model ids come from https://platform.kimi.ai/docs/models (retrieved 2026-09-29) and match the chat request schema in https://platform.kimi.ai/docs/openapi.json
Auth-failure shapethe same unauthenticated GET returned HTTP 401 with body {"error":{"message":"Incorrect API key provided","type":"incorrect_api_key_error"}} (no error.code field); the errors page lists incorrect_api_key_error — "Incorrect API key provided" under 401. No key was used
Billingthe limits page says "you need to recharge at least $1 to start using" — see Billing, 2026-09-29
Tools / structured outputdocumented on the tool-calls and response_format guides linked above; the tool-calling and structured-output paths were not exercised live, and no combined tools+schema request was sent — structuredOutputWithTools stays false
Live capability sweepnot run. evidence.liveMatrix is null. Before treating this provider as production-ready, run npx tsx test/continuous-test-suite-provider-matrix.ts --provider=moonshot-ai with a real key and record the result.

Do not treat this entry as equivalent to a live-verified Tier-2 provider (e.g. FriendliAI, Novita AI) until that live matrix has been run and evidence.liveMatrix is filled in.

Troubleshooting​

The rows restate what the vendor's errors page or Troubleshooting page says.

SymptomCause (vendor wording)Fix (vendor wording)
HTTP 401 incorrect_api_key_error"The API key was not provided, or the key is incorrect.""Make sure the endpoint matches the platform where the key was created."
HTTP 404 resource_not_found_error"Model not found, or this account does not have permission to access the model""Check the spelling of the model parameter and the account tier."
HTTP 429 engine_overloaded_error"The service node is under high load (for example, peak-hour capacity pressure).""Wait as indicated by Retry-After, reduce concurrency, and retry with exponential backoff."
HTTP 429 rate_limit_reached_error"an organization-level concurrency, RPM, TPM, or TPD limit was reached""Reduce request frequency, or see Top-up and Rate Limits to upgrade your tier"
HTTP 429 exceeded_current_quota_error"insufficient balance, an overdue account, or an expired voucher""Check available_balance with the balance API and top up."
HTTP 504 on a long non-streaming request"No response from the server for 900 seconds; the gateway returns an HTML timeout page.""Use streaming output (stream: true)."
finish_reason is length"the number of tokens generated by the current model exceeded the max_completion_tokens parameter""we recommend increasing max_completion_tokens appropriately"
Error after passing temperature or top_p"passing any other value returns an error""do not pass it explicitly"
Error after passing tool_choice: required to K2.x"These models do not support required and return an error if it is passed""only kimi-k3 supports it"

See also​