This is a 50% REAP of lossytensors based on huihui's derestriction of GLM 4.5 Air, uploaded by concedo.
Since the previous (failed-ish) attempt the combined calibration dataset was generalized (while still having a 60%/40% creative bias) and re-wrapped on source basis. Activations were also recorded with higher length (8192 vs 3072). As before, a sparse observer was utilized instead of materializing the entire [128, dim, tok] block when only top_k ones are actually routed.