MagicQuant (v2.0) - Hybrid Mixed GGUF Models + Unsloth Dynamic Learned Quant Configurations + Benchmark table with collapsed winners and more

I spent the past 5+ months building a pipeline that creates hybrid GGUF quant mixes. I also built it to learn from Unsloth (or other) models by utilizing their quant to tensor assignment. And some architectures like Qwen3.6 27B have super weird patterns that can get genuinely lower KLD while droppin