inference/research
verifiedhigh · 0.90updated 2026-10-03

DeepSpeed-MoE: Advancing Mixture-of-Experts Inference and Training to Power Next-Generation AI Scale

paper

Fields

id
sources/flop-deepspeed-moe
URL
https://arxiv.org/abs/2201.05596
Publisher
arXiv (DeepSpeed team / Microsoft)
Kind
paper
Published
2022-01-14
Accessed
2026-10-03
notes
Reports up to 3.7x MoE model-size reduction from expert pruning/merging and a highly optimized inference system giving 7.3x better latency and cost than existing MoE inference solutions.

Notes

Reports up to 3.7x MoE model-size reduction from expert pruning/merging and a highly optimized inference system giving 7.3x better latency and cost than existing MoE inference solutions.

Cross-references

This record references no other records.