Modal ships @modal.clustered for multi-node training/inference with InfiniBand/RDMA up to 6.4 Tbps — lowers the ops tax for teams that need burst multi-node without owning the fabric.
Modal declared multi-node Clusters generally available on 1 Oct 2026 via @modal.clustered, with automated InfiniBand/RDMA setup (claimed up to 6.4 Tbps) for PyTorch/NCCL workloads that previously needed hand-tuned fabric. The pitch is burst multi-node without owning cluster ops — spin up, train or serve, tear down. Useful for teams that hit single-node walls on training or high-concurrency inference; still treat advertised bandwidth as a product ceiling, and check plan GPU limits before assuming you can scale a full research fleet overnight.
Source: Modal — Clusters generally available (2026-10-01)