ResBM slashes pipeline bandwidth 128×

This looks more like a training-systems breakthrough than a new model family: if the results hold, it attacks one of the hardest constraints in scaling across weak or decentralized links.

• –The explicit identity path is the key design choice, because it tries to preserve optimization behavior while compressing inter-stage communication.
• –The comparison point is Subspace Models, but ResBM’s pitch is cleaner because it is trainable end-to-end as part of the architecture rather than relying on a more constrained optimization scheme.
• –The fact that the strongest compressed runs use Muon suggests optimizer choice still matters, so the headline gain is not purely architectural.
• –If the compute and memory overhead really stay low, this could make pipeline parallelism more practical for heterogeneous clusters, edge setups, and “internet-grade” training networks.