import litellm

This claims frequency_penalty is supported:

params = litellm.get_supported_openai_params(model='gemini/gemini-2.5-flash') print('frequency_penalty' in params) # True

But this fails:

litellm.drop_params = True # has no effect since param is "supported" response = await litellm.acompletion( model="gemini/gemini-2.5-flash", messages=[{"role": "user", "content": "Hello"}], temperature=0, max_tokens=50, frequency_penalty=0.4, )

GeminiException BadRequestError - {"error": {"code": 400, "message": "Penalty is not enabled for models/gemini-2.5-flash"}}

TL;DR

The litellm library incorrectly lists frequency_penalty as a supported parameter for the Gemini model, causing a 400 error when the parameter is passed to the Gemini API.

Guidance

• Verify the supported parameters for the Gemini model using the Gemini API documentation to ensure accuracy.
• Consider modifying the get_supported_openai_params function to exclude frequency_penalty for Gemini models that do not support it.
• If using drop_params=True is not effective, try removing the frequency_penalty parameter from the request to the Gemini API.
• Check if there are any updates or patches available for the litellm library that address this issue.

Example

# Example of removing frequency_penalty parameter
response = await litellm.acompletion(
    model="gemini/gemini-2.5-flash",
    messages=[{"role": "user", "content": "Hello"}],
    temperature=0,
    max_tokens=50,
    # Remove frequency_penalty parameter
)

Notes

The issue seems to be specific to the litellm library and its interaction with the Gemini API. The solution may require updates to the library or modifications to the code that uses it.

Recommendation

Apply workaround: Remove the frequency_penalty parameter from the request to the Gemini API, as shown in the example above, until a fix is available for the litellm library.