Abliterated version of the new Qwen3.6-35B-A3B up on HF
Pushed an abliterated Qwen3.6-35B-A3B to HF. Worth noting because MoE abliteration is genuinely different from dense — the refusal signal lives in the expert path, not attention, so standard Q/K/V LoRA doesn’t cut it. Approach (Abliterix framework): L…