User Alleges Qwen3.8-27B Quietly Weakened

A LocalLLaMA user reports that a Q4 quant of Qwen3.8-27B, run via llama.cpp on an RTX 4090, once handled complex tasks well but now fails basics and forgets instructions. The user suspects a secret quantization nerf to make Qwen4 look better, offering no evidence.

• Developer Measures Output Damage From Engine Speed Flags2026-11-01
• Developer runs Qwen 3.8 Flash Next on an RTX 3060 at 22 tok/s2026-10-31
• User runs 748B DeepSeek-V4.1-Flash on a €2k PC2026-10-22
• Community tests full n-gram table for Qwen 3.8 Flash Next2026-10-18
• Qwen 3.8 Flash runs at ~21 tok/s on RTX 30602026-10-10