Is this model expected to be able to use tools? Llama 3.3 70B can, so I'd expect this to be able to. I wonder if the tokenizer config forcing the model to start generation with <think>\n is interrupting its ability to output json and make a tool call.
