← All models

Kimi-K2-Instruct

Published Publisher Moonshot AIkimi-k2-0711-preview

Positioning and Release Date

Kimi-K2-Instruct is a post-trained instruction model in the Kimi K2 family, intended for general conversation and agent applications, with a focus on tool calling and task execution rather than a long-thinking mode. The API version kimi-k2-0711-preview was released on July 11, 2025. (kimi.com)

Capabilities and Uses

  • Uses a mixture-of-experts (MoE) architecture with approximately 1 trillion total parameters, of which approximately 32 billion are activated per token. The 0711 version has a context window of 128K tokens; do not confuse it with the 256K context window of the later 0905 version. (github.com)
  • Suitable for knowledge question answering, mathematics and coding tasks, as well as data analysis, web development, and software development workflows involving tool calls. Actual tool execution depends on the application providing the required environment and permissions. (kimi.com)
  • The model code and weights have been officially released under the Modified MIT License. This license imposes additional display requirements on commercial products or services that reach the specified scale; consult the original license before use. (github.com)

Limitations and Service Status

  • The original Kimi K2 does not support visual input and is not a long-thinking model. Difficult reasoning tasks or unclear tool definitions may cause excessive token generation, truncated output, or incomplete tool calls; enabling tools may also reduce performance on some tasks. (kimi.com)
  • As of November 18, 2026, both the official domestic and international model lists list kimi-k2-0711-preview as discontinued. The weights can be self-deployed, but this does not mean that the official API remains available. (platform.kimi.com)

API Pricing

Query date: November 18, 2026. The official domestic pricing page and official international pricing page were checked separately. The current pages on neither site could verify the RMB or USD input and output unit prices for this discontinued ID; the prices are therefore unverified, and prices for other versions are not substituted. According to the official documentation, model inference is generally billed separately by input and output token usage. (platform.kimi.com)

Official Sources