Qwen 3.6-27B FP8 faces JSON truncation in vLLM

Qwen 3.6's "Thinking Preservation" architecture is a major leap for reasoning but creates friction for existing parser state machines. The interleaved thought blocks are breaking tool-call serialization, particularly when the model attempts to generate long-form content like JavaScript files.

• –The `qwen3_xml` and `qwen3_parser` implementations in vLLM currently struggle to close tags correctly, leading to "unending" thought blocks that consume context and truncate output.
• –This "parser tax" highlights the fragility of relying on simple JSON parsers as open-weight models move toward proprietary reasoning formats.
• –Temporary fixes involve switching to the `hermes` parser or using Unsloth's optimized GGUFs, which handle nested object parsing more robustly.
• –The failure emphasizes a growing bottleneck in inference engines that must now handle high-frequency internal reasoning loops without breaking the communication layer.