Change default model for InferenceClientModel to fix ToolCallingAgent incompatibility
Context:
The ToolCallingAgent is currently incompatible with the default model specified in InferenceClientModel, which is Qwen/Qwen2.5-Coder-32B-Instruct. This breaks the out-of-the-box experience for users trying to use tool-calling with the default settings.
Problem Details:
When using the default configuration, two issues arise:
A 422 Unprocessable Entity error is returned because the Qwen model endpoint does not support the default tool_choice="required" parameter.
Even after patching the agent to send tool_choice="auto", the server returns a 500 Internal Server Error. This indicates a deeper incompatibility with the model's tool-calling implementation on the backend.
While this backend issue with the Qwen model should be addressed, the smolagents library can provide a working default experience in the meantime.
Proposed Solution:
Extensive testing has shown that other models, such as mistralai/Mixtral-8x7B-Instruct-v0.1, work flawlessly with the ToolCallingAgent without any code modifications.
This PR changes the default model_id in InferenceClientModel from "Qwen/Qwen2.5-Coder-32B-Instruct" to "mistralai/Mixtral-8x7B-Instruct-v0.1".
This is a pragmatic solution that ensures new users have a functional agent by default, avoiding the need for them to debug server-side issues or immediately override the model choice.