Calling DeepSeek's R1 model returns an empty message

[Mark531]

Calling DeepSeek's R1 model with this prompt returns an empty message. Since this prompts works perfectly when I use DeepSeek's chat, it makes me think of a bug on OpenRouter's side.

from langchain_openai import ChatOpenAI

free_key = "<KEY FOR FREE R1 MODEL>"

prompt = """
Travelers in different European cities are looking for a destination city in which to meet, knowing that their choice of city will minimize the carbon footprint of all their travel.

They can travel by train (up to 4 hours) or by plane.

The destination city can be anywhere in Europe, and must have a population of at least 10,000.

Here are the departure cities:
Lyon
Zurich
Venice

Which 10 European destinations would you recommend for their reunion, with a description of the journey and the carbon footprint (in kg of CO2) of each trip?
"""

model = ChatOpenAI(
  base_url="https://openrouter.ai/api/v1",
  model="deepseek/deepseek-r1:free",
  api_key=free_key,
  max_tokens=32000,
  temperature=0
)

result = model.invoke(prompt)

Is it a consequence of the limitations of the free model or is it a bug?

[louisgv]

Hey there -- you will want to call OpenRouter with the include_reasoning in the extra_body to get the reasoning.

[Mark531]

Hi @louisgv well, I'm not interested in the reasoning part, just in the content. And in that case it is empty. Did you try it?

[Mark531]

I have a lead! I noticed that in the output, the content field was empty as I said, but the response_metadata field was full of tokens' logprobs... So, I deactivated the outputting of logprobs:
model = model.bind(logprobs=False)
And now I could get a proper response in the content! I still don't understand why both fields couldn't be filled (even though I'm not interested in the logprobs)...