Microsoft presents ZeRO
Presented ZeRO — partitioning optimizer, gradient and parameter states opened a path to training at 100B+, with models up to 13B possible without model parallelism
Why it matters
Removed the memory barrier to 100B-class training
Source — Primary sources first; secondary reports are labelled “Press”
Paperarxiv.orghttps://arxiv.org/abs/1910.02054Added in Week 35, 2026 · See that week