Comment by impossiblefork

I mean, the model architecture is basically that of GLM 5.3 with slightly more parameters and slightly more active parameters, so it should probably be able to beat GLM 5.3 enough post-training.