Any attempt to deal with "<think>" in the code gets it replaced with "<tool_call>".
Both in inference.cerebras.ai chat and API.
Same model on chat.qwen.ai doesn't do it.