Skip to main content

​
Using the Kimi K2 family in agentic workflows

​
Always set max_tokens

Models in the Kimi K2 family can produce very long reasoning traces before arriving at a final answer. In agentic workflows where output is parsed and passed to downstream steps, always set max_tokens explicitly: 
response = client.chat.completions.create(
    model="accounts/fireworks/models/kimi-k2-instruct",
    messages=messages,
    max_tokens=512,
    tools=tools,
)
Starting points by output type:
Output typeSuggested max_tokens
Tool call responses256–512
Short text512–1024
Structured JSON1024–2048
Long-form reasoning4096+

​
Tool schema design

Kimi K2 family models perform best when tools have clearly distinct names, descriptions, and parameter schemas. When tools have overlapping surface areas the model may select the wrong one. 
# Less clear — overlapping descriptions
tools = [
    {"type": "function", "function": {"name": "read", "description": "Read data from a source", "parameters": {"type": "object", "properties": {}}}},
    {"type": "function", "function": {"name": "exec", "description": "Execute an operation on a source", "parameters": {"type": "object", "properties": {}}}},
]

# More clear — distinct names and explicit scope
tools = [
    {
        "type": "function",
        "function": {
            "name": "read_file",
            "description": "Read the contents of a file at a given path. Use this to inspect existing content before making changes. Do not use this to run code.",
            "parameters": {"type": "object", "properties": {}},
        },
    },
    {
        "type": "function",
        "function": {
            "name": "execute_command",
            "description": "Run a shell command and return its output. Use this to run scripts, tests, or system operations. Do not use this to read file contents.",
            "parameters": {"type": "object", "properties": {}},
        },
    },
]
Best practices:
  • Name tools by their primary action, not their domain (read_file not file_tool).
  • Write descriptions that distinguish tools from each other, including what each tool is not for.
  • Avoid optional parameters that make two tools look identical with only a flag difference.

​
Timeouts for agentic loops

Inference for Kimi K2 family models can be slow on large inputs. For multi-step agents, set your client read timeout to at least 10–30 minutes per call. See Reliability and error handling.