Unable to pass additional parameters to tokenizer through the MLXModel module. For example, QWEN3 has enable_thinking flag which you can pass as a boolean value. However MLXModel does not have a provision for that to be passed.

Above should be allowed. Other examples are Cohere RAG mode in the chat template. There is currently no way to pass those parameters that I was able to find.

Example below is where the value is passed but it is ignored as it is a tokenizer parameter rather than a model one.

from smolagents import MLXModel
model = MLXModel(model_id="mlx-community/Qwen3-8B-4bit", max_tokens=128)

messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Explain quantum mechanics in simple terms."}
        ]
    }
]
response = model(messages, {"enable_thinking":False})
print(response)