“By 2026, post-training recipes like that of MiMo Flash V2 fragmented into training many specialist models which are then merged into a single generalist model.”