OPEN-1B Claims Bitwise-Reproducible Training Across Heterogeneous Commodity Hardware

arXiv 2609.17380·high signalXBlueskyLinkedInCopy link

OPEN-1B (arXiv 2609.17380, submitted 15 Sep 2026) proposes a 'fully auditable' transparency tier above open weights, arguing that no released open model is provably reproducible because floating-point non-associativity means deterministic modes do not carry across hardware. By imposing a definite order on GPU kernel reductions, data batch ordering across a data-parallel cluster, and inter/intra-node collective communication, any individual training step can be replayed on a single commodity machine and checked bitwise against the published trajectory. Since replaying a whole run on one machine is infeasible, they use a collective verification scheme where many independent auditors each certify individual steps, and they release the model, its full pretraining dataset, and every intermediate checkpoint.