Qwen3.6-27B vLLM tool calls stall

This looks like a deployment friction problem, not a model-quality one: the model can be strong, but the local serving stack still needs careful parser/template alignment to avoid empty or truncated tool calls.

• –The poster is already using the common vLLM knobs: `qwen3_coder` tool parsing, `qwen3` reasoning parsing, enhanced chat template, and thinking enabled
• –Community replies point to vLLM versioning and output-field differences, with one suggestion to verify whether reasoning is emitted as `reasoning` vs `reasoning_content`
• –Related Qwen/vLLM discussion threads report empty tool calls and agent-loop termination, which matches this failure mode more than a pure FP8 accuracy issue
• –The practical takeaway is that Qwen3.6 tool use still appears sensitive to the exact serving recipe, especially around chat templates and reasoning extraction
• –For local developers, this is a reminder that “works in the model card” and “works reliably in an agent loop” are still very different bars