Table from https://unsloth.ai/docs/models/gemma-4/qat#qat-analysis I heard that MoE models are usually more susceptible to quantization error, but what happened with the 12B? I thought lower-parameter models usually quantized worse and yet, E2B/E4B are pretty much perfect while the 12B deviates from FP16 the most. Do we have an explanation for that, or did maybe something go wrong during quantization-aware training on Google's side with the 12B in particular? I'd also be interested in the e...