SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD

SLAI

Research official 2 src. ~1 min

SLAI T-Rex is an end-to-end optimization framework for training trillion-parameter MoE models on Ascend NPU hardware, achieving 34.22% Model FLOPs Utilization — a 2.93x improvement over baseline — while maintaining large-scale training stability.

Why it matters

Ranked #1 trending on HuggingFace Daily Papers (>=100 upvotes); shows a practical path to training massive models on non-GPU hardware.

Importance: 4/5

#1 trending HF Daily Paper with >=100 upvotes.

Sources