Qwen3.6 abliterated variant lands on HF

This is a useful reminder that “uncensoring” MoE models is not just a bigger LoRA problem; refusal behavior can live in expert routing and expert-specific projections.

• –The method targets O-proj, MLP down-proj, and expert slices while explicitly disabling Q/K/V, which matches the claim that the safety signal is routed through MoE experts.
• –The 7/100 refusal result is more credible than many flashy abliterated model cards because it uses longer generations and a judge model instead of simple keyword checks.
• –Router biasing toward selected “safety experts” is a strong intervention, but it also raises the risk of brittle behavior outside the exact eval set.
• –Low KL divergence suggests the base model’s general behavior is being preserved reasonably well, which matters more than raw uncensoring if the model still needs to be usable.
• –This sits in the same broader trend as other Qwen abliterations: community demand is clearly for local, less-filtered variants, especially on MoE checkpoints where the intervention surface is different.