kimi-k3

Start with low for extraction, rewriting, and short questions; try high for analysis with several steps, and compare max on difficult reasoning or coding tasks. reasoning_effort defaults to max and offers only low, high, and max, without medium or none. Higher effort may generate more reasoning tokens and take longer, so compare answer quality and usage on your task.

Thinking is always enabled for this Kimi K3 integration. Try low to reduce work instead of sending thinking.type=disabled. The Playground omits thinking, temperature, and top_p. When migrating from another model, remove those controls and select an effort using the top-level reasoning_effort field.

Generated reasoning is also billed as output, so the visible answer length is not the total output count. usage.completion_tokens already includes reasoning tokens; completion_tokens_details.reasoning_tokens, when returned, is a breakdown rather than an extra charge to add. Calculate ordinary input, cache reads, and output separately: divide each token count by one million and multiply by its current USD rate shown on this page.

Keep the required conversation history in order and replay the complete assistant message returned by the API, including content, any reasoning_content, and any tool_calls. Saving only the final text discards context used for reasoning or tool continuation. This Playground preserves the reasoning content returned by the model.

No. The model proposes a function and its arguments; your application executes it. Append the original assistant message and a role=tool result message to messages. Its tool_call_id must match the original call id, then request the next response. Do not send the tool result only as an ordinary user message.

response_format.type=json_object requests JSON, while your application still validates its business fields. To constrain fields and types, use response_format.type=json_schema with response_format.json_schema.schema and response_format.json_schema.strict=true. Parse the final JSON from choices[0].message.content, without concatenating reasoning_content. The JSON output switch in this Playground uses json_object.

This HiAPI Kimi K3 integration currently exposes text input and text output only. The official model's visual capabilities do not mean image or video input is enabled here. Use text messages rather than copying image_url or video_url examples from the provider. Any future input support will be reflected in this page's capability list and Playground controls.

This integration currently exposes POST /v1/chat/completions with model=kimi-k3; Responses is not an enabled endpoint for it. With the OpenAI SDK, use chat.completions.create and messages, update the reasoning fields, and preserve complete assistant history. You can reuse one HiAPI API key, but existing Responses code also needs its request and response handling adapted.

Cache-read billing uses the actual usage.prompt_tokens_details.cached_tokens returned for the request. Repeated text does not guarantee a hit. Keep the initial system prompt, tool definitions, and fixed context stable, then append new messages; changing reasoning_effort can also affect cache hits. When cached_tokens is zero, do not estimate that input at the cache-read rate.

No. Context size and the maximum output for one request are different limits. This page does not infer an output limit from the 1M context window and displays “-” for an unconfirmed maximum. Set max_tokens as the output budget for a request. If finish_reason is length, check whether an output limit was reached before splitting the task or continuing the response.