Microsoft Research open-sources DeepSpeed
Open-sourced DeepSpeed — claimed training of 100 billion+ parameter models, up to 5 times the throughput, and memory savings of up to 64x on 64 GPUs
Why it matters
The layer that made extremely large-scale training possible without rewriting code for model parallelism
Source — Primary sources first; secondary reports are labelled “Press”
Officialmicrosoft.comhttps://www.microsoft.com/en-us/research/blog/zero-deepspeed-new-system-optimizations-enable-training-models-with-over-100-billion-parameters/Added in Week 35, 2026 · See that week