Stepfun

Stepfun

What this is

Stepfun provides a suite of large-scale language models including Step-1 (text), Step-1V (multimodal/vision), and Step-2. It offers an OpenAI-compatible API interface, allowing developers to integrate high-performance Chinese-centric LLMs with context windows up to 256k.

Installation

Stepfun uses the standard OpenAI SDK for integration:

pip install openai

Key concepts

• API Compatibility: Uses the OpenAI client but requires a custom base_url.
• Model Tiers: Models are categorized by context window size (e.g., step-1-8k, step-1-32k, step-1-128k, step-1-256k).
• Step-1V: A multimodal model capable of understanding images and text simultaneously.
• Step-2: Next-generation model with improved reasoning and knowledge capabilities.

Correct usage patterns

Basic Chat Completion

Always set the base_url to https://api.stepfun.com/v1.

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_STEPFUN_API_KEY",
    base_url="https://api.stepfun.com/v1"
)

response = client.chat.completions.create(
    model="step-1-8k",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Explain quantum entanglement in Chinese."}
    ]
)
print(response.choices[0].message.content)

Vision Input (Step-1V)

Use the step-1v-32k model for image-related tasks.

response = client.chat.completions.create(
    model="step-1v-32k",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "What is in this image?"},
                {
                    "type": "image_url",
                    "image_url": {"url": "https://example.com/image.jpg"}
                }
            ]
        }
    ]
)

Common mistakes to avoid

• Incorrect Base URL: Forgetting to set base_url="https://api.stepfun.com/v1" will cause the SDK to default to OpenAI's servers.
• Model Naming: Using OpenAI model names (like gpt-4) instead of Stepfun names (like step-1-256k).
• Context Overload: Choosing a model with a context window too small for the input (e.g., using 8k for a 100k document).
• Vision Model Selection: Trying to perform image analysis with step-1-8k instead of the multimodal step-1v series.

File and folder conventions

• Environment Variables: Store the API key in .env as STEPFUN_API_KEY.
• Configuration: Centralize client creation in a client.py or config.py file to ensure the base_url is consistently applied.