Kimi K2¶

Development line: project:kimi-k2 · thread kimi-k2-development
Last event: 2025-07-13 · 1 dated since 2025-07-13 · Researched: 2026-09-04 · confidence: medium

What it is¶

Kimi K2 — an open-weight MoE language-model family for builders who need a self-hosted text-and-tool-use model rather than the current Kimi hosted API.

• Kimi-K2-Base for research and fine-tuning.
• Kimi-K2-Instruct for chat and agent loops.
• Local OpenAI-compatible chat and tool calling; the former hosted API also offered Anthropic compatibility.

Development line¶

• 2025-07-13 — Kimi K2 public availability references. On 2025-07-13, official project, source-code, model-publisher, web-chat, and mobile-access links recorded Kimi K2. These dated references mark a public-availability milestone for the project. The exact release wording and version details remain to be verified from source research.

What changed¶

• 2025-07-13 — Kimi K2 Base and Instruct were already public. Official records date the initial model and kimi-k2-0711-preview API release to 2025-07-11, so 13 July is not a distinct K2 version.
• 2025-08-01 — kimi-k2-turbo-preview added a high-speed hosted option, with a time-limited 50% launch promotion.

How to use this¶

From 2025-07-13, evaluate Kimi K2 from official project and source references instead of third-party summaries. Confirm developer-platform model availability separately before integrating an API.

1. Choose the route first: retain K2 only for an existing self-hosted or reproducible deployment. For a new hosted integration, choose a currently listed Kimi model, because all kimi-k2 API IDs are discontinued. — https://platform.kimi.ai/docs/models
2. For local chat or agents, start from the official Kimi-K2-Instruct checkpoint rather than Base. Base is the foundation checkpoint for research and custom post-training. — https://github.com/MoonshotAI/Kimi-K2
3. Serve the K2 weights with one of the officially recommended engines: vLLM, SGLang, KTransformers, or TensorRT-LLM. — https://huggingface.co/moonshotai/Kimi-K2-Instruct
4. Use the local chat-completions interface and start Kimi-K2-Instruct at temperature 0.6. Keep a system prompt when no special instruction is needed. — https://huggingface.co/moonshotai/Kimi-K2-Instruct
5. For a tool loop, send the available tool schemas, execute the returned call in your application, append its result, and continue until the completion no longer requests tools. — https://github.com/MoonshotAI/Kimi-K2
6. For a new Kimi API call, create an API key, use the current model ID from the model list, and call POST /v1/chat/completions at https://api.moonshot.ai/v1. — https://platform.kimi.ai/docs/api/chat

Best practices¶

• State the task, relevant context, output format, and target length explicitly. The prompt guide warns that the model cannot infer unstated requirements. — https://platform.kimi.ai/docs/guide/prompt-best-practice
• Separate supplied material with delimiters and spell out ordered steps for multi-stage work. — https://platform.kimi.ai/docs/guide/prompt-best-practice
• Use representative examples when output style or schema matters. Require answers to stay within supplied reference material when grounding is required. — https://platform.kimi.ai/docs/guide/prompt-best-practice
• Do not treat tool calling as one request: retain the returned assistant tool-call message and append each tool result before asking the model to continue. — https://huggingface.co/moonshotai/Kimi-K2-Instruct
• Do not hard-code historical K2 preview names for hosted use. They are discontinued and unsupported, so migrate and recheck the current model list. — https://platform.kimi.ai/docs/models
• Budget API traffic at the user level, not the key level: concurrency, RPM, TPM, and TPD limits are shared across models. — https://platform.moonshot.ai/docs/introduction#text-generation-model

Superseded by this¶

• 2025-09-05 — kimi-k2-0711-preview as the current hosted K2 model for new integrations; superseded by kimi-k2-0905-preview.
• 2026-05-25 — any kimi-k2 preview, Turbo, or Thinking identifier as a supported hosted API target; the K2 series was discontinued.
• 2026-08-31 — kimi-k2.5 as a supported API fallback for K2-era integrations; it was retired.

Still unknown¶

• The source links establish K2's public artifacts and releases, but not the exact claims made in the 2025-07-13 and 2025-08-01 messages whose text is unavailable.
• First-party dates conflict. The platform changelog, the 0905 announcement, and the research catalog place the original K2 release on 2025-07-11, while the current Kimi Agent overview says 2025-09-05. The latter may mean the 0905 update, but it is not explicitly corrected there.
• The July 2025 Kimi app links and the K2 model release are related product surfaces, but they do not prove which model version powered each app client on that day.
• Official sources retire K2 from the hosted API, but do not state a post-retirement maintenance or security-support policy for self-hosted K2 weights.

Sources¶

Agent brief¶

• Subject: project:kimi-k2, thread kimi-k2-development, 1 dated events 2025-07-13 → 2025-07-13.
• Practical note: From 2025-07-13, evaluate Kimi K2 from official project and source references instead of third-party summaries. Confirm developer-platform model availability separately before integrating an API.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.