Psyche Distributed Training Outperforms Centralized

This result matters because it suggests Psyche is not just a cheaper or more open way to train models, it is already good enough to produce a better model than a normal centralized run on the same 36B base. Hermes 4.3 36B was post trained entirely on Psyche, released alongside a centralized comparison, and Nous reported the Psyche version ahead on its downstream evaluations, while also extending context to 512K. That turns decentralized training from a systems experiment into evidence of product grade model quality.1 2 3 4

• The key comparison is unusually clean. Both releases sit on the same ByteDance Seed 36B base, so the gap is not coming from a bigger model. It points to the post training recipe, data flow, and training dynamics on Psyche as the source of the lift.1 2
• Psyche is built for messy internet scale training rather than a tightly controlled data center cluster. That matters because uneven GPUs and flaky connections normally make synchronous training break down. Beating a centralized run despite that overhead shows the system is handling real world heterogeneity rather than avoiding it.4 5
• Atropos helps explain why this can matter on evals. It is the environment layer for collecting reinforcement learning trajectories across many tasks, so Nous can generate training experience, score model behavior, and feed those examples back into post training. Better environments can improve the quality of what the distributed network is learning from.6 7

The next step is turning this from one strong release into a repeatable advantage. If Nous keeps showing that Psyche trained models match or beat centrally trained peers, the company can use distributed training as both a model quality engine and a supply advantage, then monetize the resulting agents and hosted usage through Portal.3 4 8