Why is the temperature of reasonning models hardcoded to 1 ?

[jeremycochoy]

In query.py we have:

    # perform temperature sampling if list provided                             
    # set temperature to 1.0 for reasoning models                               
    if kwargs_dict["model_name"] in (
        REASONING_OAI_MODELS
        + REASONING_CLAUDE_MODELS
        + REASONING_DEEPSEEK_MODELS                                           
        + REASONING_GEMINI_MODELS
        + REASONING_AZURE_MODELS
        + REASONING_BEDROCK_MODELS
    ):
        kwargs_dict["temperature"] = 1.0
    else:
        kwargs_dict["temperature"] = random.choice(temperatures)

What is the rational behind the temperature being hardcoded to 1? DeepSeek's documentation recommend a temperature of 0 for coding, and I experienced some issues with following instructions that was resolved by lowering the temperature to 0.7.

[RobertTLange]

Hi @jeremycochoy -- most reasoning models only support a temperature of 1 (probably due to post-training). E.g. OAI, Anthropic models. At some point this was the same for R1. I will double check this and adapt.

[jeremycochoy]

Ok. Qwen3 (14B quantized that we run locally) and Deepseek-reasoner (official api endpoint) support lower temperatures. Also, the pricing of deepseek changed (its now more affordable, 0.48?/M tokens).