Mistral support through `OpenAIChatCompletionClient`.

[ekzhu]

Confirmation

• I confirm that I am a maintainer and so can use this template. If I am not, I understand this issue will be closed and I will be asked to use a different template.

Issue body

Mistral AI models can be added to the list here: https://github.com/microsoft/autogen/blob/fbdd89b46bf3883efe8b0e0838370b9d74561ee7/python/packages/autogen-ext/src/autogen_ext/models/openai/_model_info.py

And the base_url should automatically be set if the model matches a Mistral AI model. API key should be detected from the environment. Use the support for Gemini model as a reference.

PR should address: #6147

[SongChiYoung]

Check : #6158

👋 While working on supporting Mistral models (e.g., mistral-large-latest, codestral-latest), I noticed that the current approach for inferring model_family via prefix matching may not work reliably anymore.

This is the current implementation:

def _find_model_family(api: str, model: str) -> str:
    for family in MESSAGE_TRANSFORMERS[api].keys():
        if model.startswith(family):
            return family
    return "default"

This logic worked well for cleanly prefixed models like gpt-4 or claude-3, and I originally wrote it when I was still getting familiar with AutoGen’s structure.
However, newer model names like mistral-large-latest, codestral-latest (but same mistral family, if someone want) introduce ambiguity, making prefix-based family inference fragile.

⸻

✅ I’m preparing a PR that refactors this logic in the following way:
• Allow model_family to be passed explicitly
• Fallback to _find_model_family() only when the value is "unknown"
• Integrate with the existing ModelInfo and ModelFamily infrastructure for consistency and clarity

Example usage:

        oai_messages_nested = [
            to_oai_type(
                m,
                prepend_name=self._add_name_prefixes,
                model=create_args.get("model", "unknown"),
                model_family=self._model_info["family"],
            )
            for m in messages
        ]

This provides more robust and future-proof transformer selection, especially for models with non-standard names like Mistral.

⸻

📌 I’ll open a PR soon with this change.

Would it also make sense to open a separate issue to track the broader design consideration around model family inference and registry consistency?
Happy to do that if it would help clarify the motivation and long-term direction.

Feedback welcome! 🙏

Check : #6158

[ekzhu]

Would it also make sense to open a separate issue to track the broader design consideration around model family inference and registry consistency?

Can we use #6011 as the roadmap issue?

[SongChiYoung]

Eric Zhu (@ekzhu) Cool.
So, how could I do that...? Maybe I does not have capability of adding that issue to #6011.

[ekzhu]

You can comment on that issue and we can create sub issues from your comment.

[Aboudjem]

Took a look at the Gemini implementation in _model_info.py and it maps cleanly to Mistral's API. They support the same OpenAI-compatible chat completions endpoint, so it's mostly adding model entries and auto-detecting the base URL + API key from MISTRAL_API_KEY. I can put a PR together for this.