What do you think about ToMoE v2 paper, converting dense model to MoE model at near lossless accuracy? I feel Qwen3.8-27B-A16B or something along those lines would be amazing, though there are architectural hurdles, as well as need folr training data.