Qwen3.6-27B NVFP4 and FP8 quantizations cause agent loops on Blackwell GPUs

A user reports severe reliability issues when running Qwen3.6-27B with NVIDIA's NVFP4 and official FP8 quantizations in agentic workflows, contrasting with flawless performance in the BF16 baseline.

• The model exhibits mid-task halting and failure loops on an RTX PRO 6000 Blackwell GPU using vLLM 0.24.0.
• Increasing repetition penalty fails to break the loops, causing the model to alternate between two distinct failure phrases instead of repeating one.
• NVFP4 quantization shows significantly more frequent degradation than FP8, while BF16 remains unaffected.

The author seeks confirmation on whether these issues are known artifacts of low-bit quantization for thinking models or if specific vLLM configuration adjustments can resolve them.