Kimi K2.6

Preserve reasoning with the right setting

The model card documents preserved reasoning across turns. The official API uses thinking.keep set to all; vLLM and SGLang use the preserve_thinking chat-template option. The model card recommends preservation in thinking mode.Kimi K2.6 model card

Practical guidance. A request option from a local inference example may not be the option required by the hosted API. Keep these configurations separate and check the returned conversation fields.

Use the input that shows the problem

The API guide combines media understanding with tools. Its examples let the model request a relevant video segment rather than require every segment in the initial prompt.Kimi K2.6 API guide

Practical guidance. This pattern can make a long recording easier to inspect: first describe the event to find, then supply the selected interval. Verify that the final answer refers to that interval rather than to unrelated footage.

Weights have a separate lifecycle

Practical guidance. A planned API shutdown does not remove files already published as model weights. If you choose self-hosting, evaluate the license, hardware and inference software as a separate project. Do not assume it will preserve the hosted service’s cost or operating behavior.