kimi-k3.1-pro

K3.1 Pro is post-trained from K3 for coding, agent tool workflows, and long documents, and it is called the same way: the same /v1/chat/completions endpoint, always-on reasoning, and low/high/max efforts. Start new projects on kimi-k3.1-pro. To migrate from K3, change model to kimi-k3.1-pro, then compare quality and usage on your own tasks. The two models have different rates; see the model pricing.

Try low for routine extraction, rewriting, and short answers; high for multi-step analysis; and max for hard reasoning or whole-project coding. reasoning_effort defaults to max and accepts only these three values; there is no medium or none. Higher efforts usually spend more reasoning tokens and take longer, so choose by answer quality and usage.

Reasoning is always on for K3.1 Pro, and temperature=1 and top_p=0.95 are fixed. Omit thinking, temperature, and top_p; a request with any other temperature or top_p value returns HTTP 400 and is not billed. To reduce cost, lower reasoning_effort to low first. When migrating from another model, remove these sampling parameters.

Reasoning is billed as output, so the visible answer length does not reflect total output. usage.completion_tokens already includes reasoning tokens; completion_tokens_details.reasoning_tokens is a breakdown, so do not add it twice. max_tokens also counts reasoning, and a very small value can cut the response off before the answer appears.

Send back the complete assistant message returned by the API, including content, reasoning_content, and tool_calls; keeping only the final text loses context. tool_calls are proposals that your application executes. Append a role=tool result whose tool_call_id matches the original call, then request the next turn. tool_choice accepts auto, none, and required; naming a specific function is not supported because reasoning is always on.

response_format.type=json_object returns JSON, but your application still validates the fields. To constrain fields and types, use response_format.type=json_schema with json_schema.schema and strict=true. Parse choices[0].message.content; do not include reasoning_content.

The official K3.1 Pro understands images and video, but this HiAPI integration currently exposes text input and text output through Chat Completions only. Do not copy the official image_url, video_url, or Responses examples. Any future input types will appear in this page's capabilities and Playground controls.

Repeated request prefixes are cached automatically with no extra parameter. Cached tokens appear in usage.prompt_tokens_details.cached_tokens and are billed at the cache-read rate; the rest of the input is billed as regular input. Cache hits are not guaranteed; see the model pricing for each rate.