Opus 5.5 developed a custom quantizer for Qwen3.5-0.8B from scratch using NumPy without relying on llama-quantize, imatrix, or published recipes. The approach measured signal-carrying input columns via a float model copy to minimize rounding error where it matters most, rather than against weights. Using only the BF16 file as input and stock llama.cpp formats, the resulting quantization achieves 5.9× better quality than Unsloth's same-size output.