Kimi

Kimi K3

LLM

Kimi K3 is Kimi's flagship long-context model for advanced reasoning, coding, agent workflows, native multimodal understanding, and knowledge-intensive tasks.

kimi-k3

PROVIDERSKimiENDPOINTSopenaiCAPABILITIEStext

PRICINGDynamic Pricing

Prices vary by usage tier, request parameters or context conditions.

Tier	Input / 1M	Output / 1M	Cache Read / 1M
`base`	$2.66	$13.3	$0.266

Parameter	Type	Default	Range / values	Description
`model` *	`string`	`kimi-k3`	—	Model ID used for the chat completion request
`messages` *	`array`	—	—	Conversation messages used as model input
`max_completion_tokens`	`integer`	`131072`	1 ~ 1048576 - prompt_tokens	Maximum number of tokens to generate. For Kimi K3, the practical upper bound depends on the remaining 1M-token context window.
`reasoning_effort`	`enum`	`max`	low · high · max	Controls the thinking effort level. Kimi K3 always runs in thinking mode.
`temperature`	`number`	—	0 ~ 2	Controls randomness when supported by the channel or model; lower values make outputs more deterministic.
`top_p`	`number`	—	0 ~ 1	Controls nucleus sampling when supported by the channel or model.
`stop`	`string \| array`	—	—	Optional stop sequence or sequences that end generation
`stream`	`boolean`	`false`	—	Whether to stream partial responses
`stream_options`	`object`	—	—	Streaming options, such as whether to include usage information in the stream
`tools`	`array`	—	—	Tool definitions for function calling; tool type is commonly "function"
`tool_choice`	`string \| object`	`auto`	none · auto · required	Controls whether and how the model calls tools. String values include "none", "auto", and "required"; object form may be used to force a specific function when supported.
`response_format`	`object`	—	json_object	Controls structured output format, such as {"type":"json_object"}