Qwen3.6 35B shows quantization jitters

This is the kind of small, ugly eval that matters for local LLM users: one toy prompt can expose how much behavior shifts when a sparse MoE model gets squeezed.

• –The reported failure mode is not raw benchmark loss, but answer polarity flipping under lower-bit quants
• –Qwen3.6-35B-A3B’s sparse MoE shape may make per-layer or activation-sensitive quantization more important than a simple “Q4 is good enough” rule
• –The comparison with Qwen3.6-27B suggests smaller or denser variants may be more robust for local setups
• –Developers using GGUF builds should test their actual task prompts across quant levels, not assume leaderboard quality survives compression