Reasoning

​

1. Overview

• Structured thinking: Using techniques like Chain-of-Thought to break down complex problems
• Knowledge integration: Combining domain knowledge bases with common sense reasoning capabilities
• Self-correction mechanism: Enhancing result reliability through feedback loops
• Multimodal processing: Some advanced models support mixed input of text, code, and formulas

​

2. Supported model list

• THUDM:
  • THUDM/GLM-4.1V-9B-Thinking
  • Pro/THUDM/GLM-4.1V-9B-Thinking
• tencent:
  • tencent/Hunyuan-A13B-Instruct
• MiniMaxAI：
  • MiniMaxAI/MiniMax-M1-80k
• Qwen Series:
  • Tongyi-Zhiwen/QwenLong-L1-32B
  • Qwen/Qwen3-30B-A3B
  • Qwen/Qwen3-32B
  • Qwen/Qwen3-14B
  • Qwen/Qwen3-8B
  • Qwen/Qwen3-235B-A22B
  • Qwen/QwQ-32B
• THUDM Series:
  • THUDM/GLM-Z1-32B-0414
  • THUDM/GLM-Z1-Rumination-32B-0414
  • THUDM/GLM-Z1-9B-0414
• deepseek-ai Series:
  • deepseek-ai/DeepSeek-R1
  • Pro/deepseek-ai/DeepSeek-R1
  • deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
  • deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
  • deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
  • Pro/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
• ascend-tribe:
  • ascend-tribe/pangu-pro-moe

​

3. Usage recommendations

​

3.1 API parameters

​

3.1.1 Input parameters

• Input parameters:
  • Maximum Chain-of-Thought Length (thinking_budget): The number of tokens the model uses for internal reasoning. Adjusting the thinking_budget controls the length of the chain-of-thought process.
  • Maximum Response Length (max_tokens): This is used to limit the number of tokens in the final output to the user, excluding the chain-of-thought part. Users can configure this normally to control the maximum length of the response.

• After decoupling the reasoning model’s chain-of-thought process from the response length, the output behavior will follow the following rules:
  • If the number of tokens generated during the “thinking phase” reaches the thinking_budget, the Qwen3 series reasoning model, which natively supports this parameter, will forcibly stop the chain-of-thought reasoning. Other reasoning models might continue to output the thinking content.
  • If the maximum response length exceeds the max_tokens limit or the context length exceeds the context_length restriction, the response content will be truncated. The finish_reason field in the response will be marked as length, indicating that the output was terminated due to length constraints.

​

3.1.2 Return parameters

• Return parameters:
  • reasoning_content：Reasoning chain content, at the same level as content.
  • content：Final answer content

​

3.2 DeepSeek-R1 Usage Recommendations

• Set the temperature within the range of 0.5-0.7 (0.6 is recommended) to prevent endless repetitions or incoherent outputs.
• Set the value of top_p to 0.95.
• Avoid adding a system prompt; all instructions should be contained within the user prompt
• For mathematical problems, it is advisable to include a directive in your prompt such as: “Please reason step by step, and put your final answer within \boxed.”
• When evaluating model performance, it is recommended to conduct multiple tests and average the results.

​

4. OpenAI request examples

​

4.1 Stream Mode Request

from openai import OpenAI

url = 'https://api.siliconflow.cn/v1/'
api_key = 'your api_key'

client = OpenAI(
    base_url=url,
    api_key=api_key
)

# Send a request with streaming output.
content = ""
reasoning_content=""
messages = [
    {"role": "user", "content": "Who are the legendary athletes of the Olympic Games?"}
]
response = client.chat.completions.create(
    model="Pro/deepseek-ai/DeepSeek-R1",
    messages=messages,
    stream=True,  # Enable streaming output.
    max_tokens=4096,
    extra_body={
        "thinking_budget": 1024
    }
)
# Step-by-step receiving and processing the response.
for chunk in response:
    if chunk.choices[0].delta.content:
        content += chunk.choices[0].delta.content
    if chunk.choices[0].delta.reasoning_content:
        reasoning_content += chunk.choices[0].delta.reasoning_content

# Round 2
messages.append({"role": "assistant", "content": content})
messages.append({'role': 'user', 'content': "Go on"})
response = client.chat.completions.create(
    model="Pro/deepseek-ai/DeepSeek-R1",
    messages=messages,
    stream=True
)

​

4.2 Non-Stream Mode Request

from openai import OpenAI
url = 'https://api.siliconflow.cn/v1/'
api_key = 'your api_key'

client = OpenAI(
    base_url=url,
    api_key=api_key
)

# Send a non-streaming request.
messages = [
    {"role": "user", "content": "Who are the legendary athletes of the Olympic Games?"}
]
response = client.chat.completions.create(
    model="Pro/deepseek-ai/DeepSeek-R1",
    messages=messages,
    stream=False, 
    max_tokens=4096,
    extra_body={
        "thinking_budget": 1024
    }
)
content = response.choices[0].message.content
reasoning_content = response.choices[0].message.reasoning_content

# Round 2
messages.append({"role": "assistant", "content": content})
messages.append({'role': 'user', 'content': "GO on"})
response = client.chat.completions.create(
    model="Pro/deepseek-ai/DeepSeek-R1",
    messages=messages,
    stream=False
)

​

5. Notes

• API Key: Ensure you use the correct API key for authentication.
• Stream Mode: Stream mode is suitable for scenarios where responses need to be received incrementally, while non-stream mode is suitable for scenarios where a complete response is needed at once.

​

6. Common questions

• How to obtain the API key? Please visit SiliconFlow to register and obtain the API key.
• How to handle long text? You can adjust the max_tokens parameter to control the length of the output, but please note that the maximum length is 16K.