Huihui Adds Ternary Qwen Abliteration Variant

This is more valuable as a local-model experiment than as a polished drop-in checkpoint.

• –Combines refusal ablation with ternary quantization, making quality and safety tradeoffs difficult to predict.
• –PTQ1_0 and PQ2_0 files make a 27B model unusually accessible on consumer hardware.
• –Stock llama.cpp cannot run the ternary files; developers need PrismML’s custom runtime.
• –The untouched layers, modified ablation range, and quantization changes create a useful benchmark target against both parent models.
• –Huihui-AI explicitly frames the series as testing and validation, so production use deserves caution.