[RFC]: How to use OpenAI's "Reasoning Levels" (low/medium/high) in vLLM?

[WangRongsheng]

Motivation.

Hi,

According to OpenAI's documentation, there is a feature called Reasoning Levels that allows controlling the depth of reasoning in a model response. The levels are:

Reasoning levels
You can adjust the reasoning level that suits your task across three levels:

Low: Fast responses for general dialogue.  
Medium: Balanced speed and detail.  
High: Deep and detailed analysis.  
The reasoning level can be set in the system prompts, e.g., "Reasoning: high".

I would like to know:

1. How can I use this in vLLM? Should I just include "Reasoning: high" in the system prompt, or is there another preferred way to activate it?
2. What is the default reasoning level if I don’t explicitly set it?

Proposed Change.

None

Feedback Period.

No response

CC List.

No response

Any Other Things.

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[iseesaw]

The same problem, it seems that we can only use openai_harmony to process, passing in prompt_token_ids.

Reference: cookbook: using-vllm-for-direct-sampling

[suresh-now]

You can add directly to system prompt. Reasoning: Low

[Wen1163204547]

{
"model": "gpt-oss-20b-vllm",
"messages": [{"role": "user", "content": "Let $N$ denote the numbers of ordered triples of positive integers $(a, b, c)$ such that $a, b, c \le 3^6$ and $a^3 + b^3 + c^3$ is a multiple of $3^7$. Find the remainder when $N$ is divided by $1000$."}],
"max_tokens": 32000, "stream": true, "temperature": 0.6, "reasoning_effort": "low"
}

[suresh-now]

@Wen1163204547 Thanks. Is there any tool parser that acutally works/ The model doesnt invoke a tool call

"tools": [
{
"type": "function",
"function": {
"name": "retrieve",
"description": "Retrieve documentation sections based on a search query.\nThis is a tool that retrieves documentation sections based on a search query.",
"parameters": {
"type": "object",
"additionalProperties": false,
"required": ["search_query"],
"properties": {
"search_query": {
"type": "string",
"description": "The search query."
}
}
}
}
}
]

[Wen1163204547]

@suresh-now Sorry, I don't have the environment for toolcall, so I haven't tried toolcall.

[brandonbiggs]

The following worked for me:

    response = client.responses.create(
        model=model,
        input="Tell me a story",
        instructions="You are a helfpul assistant.",
        reasoning={
            "effort": "low", # "low", "medium", or "high"
            "summary": "auto", # "auto", "concise", or "detailed"
        }, 
    )

[WangRongsheng]

{ "model": "gpt-oss-20b-vllm", "messages": [{"role": "user", "content": "Let N denote the numbers of ordered triples of positive integers ( a , b , c ) such that a , b , c ≤ 3 6 and a 3 + b 3 + c 3 is a multiple of 3 7 . Find the remainder when N is divided by 1000 ."}], "max_tokens": 32000, "stream": true, "temperature": 0.6, "reasoning_effort": "low" }

copy!