Why AI models can't agree on how to format their thinking (and why you should care)

Here's the mess: different AI models are using completely different formats for their internal reasoning tokens. Qwen and DeepSeek use `<think>...</think>`, Gemma uses `<|channel>...</channel|>`, and sometimes Gemma just... doesn't use delimiters at all and throws bare text at you. It's chaos. vLLM tried to patch this with `--reasoning-parser` flags for each model, but that's basically just the maintainers playing whack-a-mole forever. And if you're building anything that actually uses these reasoning outputs? You're stuck writing custom parsers for every single model. We already learned this lesson with chat templates—looks like we're about to repeat it.