Specifying reasoning params

[velvet frigate]

Hi, I have some questions regarding reasoning params.
Is there any point of using o3-mini-high, if I specify the reasoning effort myself?

How can I find out the max supported values for reasoning max_tokens?
How do I even know if a given model supports reasoning or not?

For example here is r1:

{
  "id": "deepseek/deepseek-r1",
  "name": "DeepSeek: R1",
  "created": 1737381095,
  "description": "DeepSeek R1 is here: Performance on par with [OpenAI o1](/openai/o1), but open-sourced and with fully open reasoning tokens. It's 671B parameters in size, with 37B active in an inference pass.\n\nFully open-source model & [technical report](https://api-docs.deepseek.com/news/news250120).\n\nMIT licensed: Distill & commercialize freely!",
  "context_length": 65536,
  "architecture": {
    "modality": "text->text",
    "tokenizer": "DeepSeek",
    "instruct_type": "deepseek-r1"
  },
  "pricing": {
    "prompt": "0.00000055",
    "completion": "0.00000219",
    "image": "0",
    "request": "0",
    "input_cache_read": "0",
    "input_cache_write": "0",
    "web_search": "0",
    "internal_reasoning": "0"
  },
  "top_provider": {
    "context_length": 65536,
    "max_completion_tokens": 8192,
    "is_moderated": false
  },
  "per_request_limits": null
}

[velvet frigate]

Anyone on this?

[wide thicket]

I don't know about o3-mini-high, but for the max_tokens for reasoning, if you don't specify a value at all, it will default to its maximum value

A token born from reasoning is treated just like any other token. If you really want to specify a value, specify the max output length supported by that provider or the models max context length

I.e

max_tokens: 128000

[velvet frigate]

Clear! But is there any information at all in the models API which tells me if a model supports reasoning or not? I couldn't find anything.

It'd be great to have the following information:

• supports_reasoning true/false
• supported_reasoning_levels: ['low','mid','high', '_max_tokens']
  or something similar

[wide thicket]

There's nothing in the api currently but I agree with you on "it would be nice to know if a model does reasoning or not"

Afaik, there's only 3 models that do reasoning

Claude sonnet 3.7:thinking
R1 (and it's distills)
And qwq (and its dirivatives)

[wide thicket]

You should create a post in #1324228782114213948

[velvet frigate]

• Gemini Thinking, o1, o3

[wide thicket]

I forgot about those

5 models, then