Also, Mistral Large 4 will probably beat GLM 5.3 once post-training is done.

Relatively similar concept, but has more total parameters, more active parameters and ordinary full attention instead of deepseek sparse attention, so it should probably be GLM 5.3-beating.