Kimi-K2-Instruct

Kimi K2 Instruct is the first post-trained K2 release. It targets direct responses and tool use, while K2 Base serves further training and the later K2 Thinking release adds extended reasoning.

Tool use was part of post-training

The K2 report describes synthesized agent tasks and reinforcement learning in both real and synthetic environments. It evaluates the resulting model without extended thinking. This distinguishes the original Instruct training target from the later long-reasoning release.Kimi K2: Open Agentic Intelligence

Use the supplied tool-call format

The model card includes multi-step tool examples and a tokenizer tool-call parser. It recommends a sampling temperature of 0.6. Later card changes document corrections to tokenization and tool serialization; the repository version matters when reproducing behavior.Kimi-K2-Instruct model card

An API retirement does not remove the checkpoint

The license permits use subject to its terms and requires Kimi K2 attribution above the stated 100-million-user or US$20-million monthly-revenue thresholds.Kimi K2 Base and Instruct license

Practical guidance. Pin the checkpoint and tokenizer together. Verify tool names, arguments and tool results in the application that executes them. A generated call is a request for an action, not proof that the action succeeded.

The later Turbo endpoint

Moonshot launched kimi-k2-turbo-preview as a faster K2 service in August 2025. It is a serving variant, not a separately documented base-model release.Kimi API: dated 2025 model release log

Specifications & access

Kimi-K2-Instruct specifications
DetailValue and conditions
Launch date
AccessDownloadable
Model identifiersmoonshotai/Kimi-K2-Instructkimi-k2-0711-preview
Context window131,072 tokens

Published checkpoint limit; this is not a promise of continued API access.

Kimi-K2-Instruct checkpoint configuration
Parameters1 trillion total; 32 billion active per token.Kimi-K2-Instruct model card
Weight licenseModified MIT.Kimi K2 Base and Instruct license
Former APIkimi-k2-0711-preview is listed as deprecated; downloadable weights remain published.Kimi model catalog

Related models

  • Revised non-thinking K2 model with a doubled context window and a stronger focus on agentic coding.

    Launch date Downloadable
  • Long-reasoning K2 model trained to alternate internal reasoning with repeated tool interactions.

    Launch date Downloadable
  • Trillion-parameter pretrained model that introduced MuonClip training to the Kimi K2 family.

    Launch date July 2025Downloadable