Kimi-K2-Thinking

Kimi K2 Thinking extends the K2 family with a model designed to reason while using tools. The published checkpoint includes quantization-aware training for native INT4 inference.

Reasoning continues between tool results

The model card describes interleaved reasoning and tool calls over long tasks. Moonshot reports runs with hundreds of sequential calls in its evaluation. That result depends on the tools, task environment and generation budget, not only on the model.Kimi-K2-Thinking model card

Quantization is part of this release

Moonshot applied quantization-aware training during post-training and published compressed INT4 weights. This distribution is the Thinking model itself, rather than an additional model family created by converting its files.Kimi-K2-Thinking model card

Turbo was a serving option

The release log announces the thinking model and its Turbo version together. They remain grouped here as a model and its service variant.Kimi API: dated 2025 model release log

The weight license adds Kimi K2 attribution for products above 100 million monthly users or US$20 million monthly revenue.Kimi K2 Thinking license

Practical guidance. Budget for reasoning tokens, tool latency and execution limits. Use an inference engine that supports the released quantization format before comparing its memory requirements.

Specifications & access

Kimi-K2-Thinking specifications
DetailValue and conditions
Launch date
AccessDownloadable
Model identifiersmoonshotai/Kimi-K2-Thinkingkimi-k2-thinkingkimi-k2-thinking-turbo
Context window262,144 tokens

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

Kimi-K2-Thinking checkpoint configuration
Parameters1 trillion total; 32 billion active per token.Kimi-K2-Thinking model card
Weight licenseModified MIT.Kimi K2 Thinking license
Former APIskimi-k2-thinking and kimi-k2-thinking-turbo are listed as deprecated.Kimi model catalog

Related models

  • Multimodal K2 successor that combines image and video understanding with coding and reasoning.

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

    Launch date Downloadable