PyTorch trains 1T-parameter MoE on EU public infrastructure

On November 12, 2026, the PyTorch Blog published a post titled "Training a 1T-Parameter MoE on EU public infrastructure with PyTorch and TorchTitan", describing the team's work on pretraining a roughly 1-trillion-parameter MoE model. The work was carried out on shared, public infrastructure located in the EU, and relied on PyTorch together with TorchTitan as the training stack. The post states that scaling pretraining for a model of this size on such shared, public infrastructure had previously been viewed as unfeasible, with unique architectural and interconnect bottlenecks cited as the reasons. It further explains that the team addressed those bottlenecks by leveraging fine-grained flexibility, which allowed the pretraining effort to proceed on the shared public infrastructure rather than being blocked by the constraints that had made similar efforts impractical in the past. As of now, only this single post is publicly available, and it offers no further details about the training configuration, the resulting model, or any subsequent progress. No additional follow-up reports have surfaced to date, leaving the current state of the effort as a first-hand account from the PyTorch team with no independent confirmation or updates beyond the original announcement.