Kimi K3

Kimi K3 is available through the OpenAI-compatible Chat Completions API. It supports text and image input, reasoning content, streaming, and function calling.

Endpoint

POST /v1/chat/completions

Model

kimi-k3

Actual availability and quota depend on your account permissions.

Basic Request

curl https://cubicspaces.cloud/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [
      { "role": "system", "content": "You are a precise and concise technical assistant." },
      { "role": "user", "content": "List three principles for designing an idempotent API." }
    ],
    "reasoning_effort": "high",
    "max_tokens": 4096,
    "stream": false
  }'

In a non-streaming response, the final answer is returned in choices[0].message.content. When reasoning content is present, it is returned in choices[0].message.reasoning_content.

Reasoning Effort

reasoning_effort accepts these values:

Value	Description
low	Shorter reasoning for simple tasks
high	Deeper reasoning for complex analysis and coding
max	Maximum reasoning effort and potentially longer output

Do not use medium or any unlisted value. Unsupported values return HTTP 400.

Image Input

Place images in the user message's content array using the OpenAI image_url format. A publicly accessible HTTPS URL is recommended.

{
  "model": "kimi-k3",
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "image_url",
          "image_url": {
            "url": "https://example.com/image.jpg"
          }
        },
        {
          "type": "text",
          "text": "Describe this image and list its three most important details."
        }
      ]
    }
  ],
  "reasoning_effort": "high",
  "max_tokens": 4096
}

Streaming

curl https://cubicspaces.cloud/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [
      { "role": "user", "content": "Analyze the risks in this database migration plan." }
    ],
    "reasoning_effort": "high",
    "stream": true,
    "stream_options": {
      "include_usage": true
    }
  }'

Streaming responses use SSE. Parse these fields separately:

• choices[0].delta.reasoning_content: reasoning chunks.
• choices[0].delta.content: final-answer chunks.
• usage: returned near the end when stream_options.include_usage is enabled.
• data: [DONE]: end-of-stream marker.

Function Calling

Tools must use the OpenAI Function Calling format. tools must be an array, and every tool must contain type: "function" and a complete function object.

{
  "model": "kimi-k3",
  "messages": [
    { "role": "user", "content": "Check the current weather in Tokyo." }
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get the weather for a city",
        "parameters": {
          "type": "object",
          "properties": {
            "city": { "type": "string" }
          },
          "required": ["city"]
        }
      }
    }
  ],
  "tool_choice": "auto",
  "reasoning_effort": "high"
}

Function names must follow these rules:

• No more than 64 characters.
• Only ASCII letters, numbers, underscores, and hyphens.
• function.parameters must be a JSON object.

After the model returns tool_calls, execute the tools and append their results to the message history:

{
  "model": "kimi-k3",
  "messages": [
    { "role": "user", "content": "Check the current weather in Tokyo." },
    {
      "role": "assistant",
      "content": null,
      "tool_calls": [
        {
          "id": "call_1",
          "type": "function",
          "function": {
            "name": "get_weather",
            "arguments": "{\"city\":\"Tokyo\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "tool_call_id": "call_1",
      "content": "Sunny, 26°C"
    }
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "parameters": {
          "type": "object",
          "properties": {
            "city": { "type": "string" }
          },
          "required": ["city"]
        }
      }
    }
  ]
}

Every role: "tool" message must use the exact tool_calls[].id returned by the preceding assistant message. Return a result for every tool call before sending the next regular message.

Python SDK

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url="https://cubicspaces.cloud/v1",
)

response = client.chat.completions.create(
    model="kimi-k3",
    messages=[
        {"role": "user", "content": "Explain optimistic locking."}
    ],
    reasoning_effort="high",
    max_tokens=4096,
)

print(response.choices[0].message.content)

Parameters

Parameter	Type	Description
model	string	Must be kimi-k3
messages	array	OpenAI Chat message array
reasoning_effort	string	low, high, or max
max_tokens	integer	Maximum output tokens; recommended for this model
max_completion_tokens	integer	Maximum output tokens; accepted by some clients
stream	boolean	Whether to stream the response
stream_options	object	Streaming options such as include_usage
tools	array	Function Calling tool definitions
tool_choice	string/object	Tool selection policy

The platform currently limits the requested output maximum to 65,536 tokens. Use a smaller practical value whenever possible.

Parameters to Omit

To avoid compatibility errors, do not add these fields to Kimi K3 requests:

• temperature
• top_p
• top_k
• presence_penalty
• frequency_penalty
• prompt_cache_key
• prompt_cache_retention
• A literal extra_body field in raw HTTP JSON
• Legacy functions or function_call

The Python SDK does not require extra_body for Kimi K3. If another client provides an advanced field named Extra Body, do not place unknown extensions in it.

Common HTTP 400 Errors

Symptom	Common cause	Resolution
Unsupported reasoning_effort	medium or a custom value was used	Use low, high, or max
tools[n].function.name error	Missing, overlong, or invalid function name	Follow the function-name rules above
function.parameters must be a JSON object	The schema was sent as a string or array	Send a valid JSON object
tool_call_id does not match	A tool result has no matching prior call	Reuse the original tool_calls[].id
All tool results must be returned first	Only some parallel tool calls received results	Append one role: "tool" message for every call
Invalid messages or tools structure	Objects and arrays were mixed, or an item is incomplete	Compare the final JSON hierarchy with the examples above
JSON parsing error	Truncated payload or incorrect quote/backslash escaping	Validate the final HTTP request body with a JSON parser

When troubleshooting, inspect the final HTTP JSON sent by the client, not only the SDK call. Proxies, plugins, and advanced client settings may add fields to the final payload.