Reasoning model best practices

StepFun offers a range of reasoning models covering text reasoning, multimodal understanding, agent, and code analysis scenarios. We recommend using step-3.7-flash as the preferred option; for pure text reasoning, you can also use step-3.5-flash.

step-3.7-flash is StepFun’s flagship multimodal reasoning model. Building on the high-speed reasoning and tool-calling capabilities of step-3.5-flash, it adds native image and video understanding and supports three levels of reasoning effort, making it well-suited for agent, code, multimodal analysis, and complex planning tasks.

step-3.5-flash is a high-speed reasoning model built for extreme efficiency. Based on a sparse Mixture-of-Experts (MoE) architecture, it carries 196B parameters but selectively activates only ~11B per token, pairing the logical depth of much larger models with low-latency inference. With a 256K context window plus solid tool-calling and multi-step agent capabilities, it is well-suited to pure-text reasoning, engineering, and automation workloads.

Reasoning effort control

Models that support three levels of reasoning effort let you tune thinking depth via a parameter. The Chat Completion API uses reasoning_effort; the Messages API uses output_config.effort.

Reasoning effort	Use cases
low	Simple Q&A, summarization, rewriting, information extraction
medium	Default recommendation, suitable for general reasoning and multi-step tasks
high	Complex reasoning, math, planning, code analysis

For complete call examples, see Step 3.7 Flash quickstart. For parameter field details, see Chat Completion API and Messages API.

Chat Completion Example

Calling reasoning models for text chat works the same way across models. The following uses step-3.5-flash, a common choice for pure-text scenarios, to build a streaming conversation.

copy

import time
from openai import OpenAI

# Set your API Key and Base URL
BASE_URL = "https://api.stepfun.ai/v1"
STEP_API_KEY = "YOUR_STEPFUN_API_KEY"

# Select Model
COMPLETION_MODEL = "step-3.5-flash"

# User Prompt
user_prompt = "How many 'r's are in the word strawberry?"

client = OpenAI(api_key=STEP_API_KEY, base_url=BASE_URL)

time_start = time.time()

try:
    response = client.chat.completions.create(
        model=COMPLETION_MODEL,
        messages=[
            {"role": "user", "content": user_prompt}
        ],
        stream=True
    )
except Exception as e:
    print("Exception occurred when requesting API:", e)
    exit(1)

print("Reasoning Process:")
try:
    for chunk in response:
        # Check for reasoning content
        if hasattr(chunk.choices[0].delta, 'reasoning') and chunk.choices[0].delta.reasoning:
             print(chunk.choices[0].delta.reasoning, end='', flush=True)
        # Check for standard content
        elif chunk.choices[0].delta.content:
             print(chunk.choices[0].delta.content, end='', flush=True)

except Exception as e:
    print("\\nError occurred while processing streaming results:", e)

time_end = time.time()
print(f"\\n\\nTotal generation time: {time_end - time_start:.2f} seconds")

For input parameter details, please refer to the Chat Completion Documentation

Obtaining Reasoning Content

When StepFun’s reasoning models handle complex problems, they include a reasoning field in the output to display the model’s thinking process. Developers can check for the existence of this field to obtain the model’s thinking information.

if chunk.choices[0].delta.reasoning:
    reasoning = chunk.choices[0].delta.reasoning
    print("Model thinking process:", reasoning)

For non-streaming scenarios, you can directly extract the reasoning field to get the model’s thinking process.

msg  = completion.choices[0].message.content
reasoning = completion.choices[0].message.reasoning

Notes

• Error Handling and Logging: A Trace ID is added to model outputs. Please include this ID when reporting any issues with reasoning behavior.