兔子API 兔子API
中

model

Moonshot Kimi

Direct answerKimi currently documents K3, K2.7 Code, and K2.6 for general multimodal work, long-running coding, and existing workloads through an OpenAI-compatible API.

Updated · Reviewed

Current official portfolio

Kimi's quickstart currently highlights Kimi K3, K2.7 Code, and K2.6 and documents text, image and video input, streaming, multi-turn conversations, tools, JSON Mode, and thinking. Exact availability remains in the official model list.

Capability differences

K3 is the general multimodal and knowledge-work entry, K2.7 Code targets long-context coding and agents, and K2.6 retains general multimodal support. Check each model page for context, output, and high-speed variant pricing.

Choosing within the family

Evaluate K3 for research, long documents, and cross-modal work; K2.7 Code for repository changes, long-running development, and tool loops; K2.6 for compatible existing applications. Long context still needs retrieval, compaction, and evidence evaluation.

API use

Kimi is compatible with OpenAI SDKs through /v1/chat/completions, a Bearer token, and messages. Upload large images or videos as files rather than resending Base64 in every turn.

Limits and evaluation

The official guide notes that generation is nondeterministic. Verify model ID, file specifications, context, thinking, tools, JSON, rate limits, and caching, and set timeout, retry, and cost bounds for long requests.

Use cases

  • Long context
  • Long-horizon coding
  • Knowledge work

API protocols

  • /v1/chat/completions

FAQ

Is the model name permanent?

No. Official aliases can move and enabled API IDs can change; pin a tested ID in production and verify availability separately in the model marketplace.

How should I choose between general multimodal and Code Kimi models?

Evaluate general multimodal models for research, long documents, and cross-modal work, and Code models for repository changes, long-running development, and tool loops. Test older models for compatibility and compare accuracy, tool success, latency, and tokens on the same workload.

Does Kimi's OpenAI SDK compatibility make every parameter portable?

No. Thinking, tools, JSON, vision, and file limits remain model-specific. Use supported file workflows for larger images or video instead of resending Base64 in every turn, and set timeouts and cost caps for long requests.

  • Moonshot KimiDirect mainland China access uses https://api.moonshot.cn/v1 and a Moonshot API key. K3, K2.6, and K2.7 use different thinking and tool parameters, while this gateway uses its own token and enabled model name.
  • OpenAI-compatible API guideThis site's OpenAI-compatible routes use a Bearer token; Responses, Chat Completions, and media endpoints have distinct objects and streaming semantics and must match the target model.
  • Choosing and calling coding modelsSelect coding models by patch correctness, test pass rate, tool success, latency, and cost on real repositories, not marketplace copy or a single benchmark.

Official sources

  1. Kimi Overview Official