Alibaba and Moonshot Target 10T Parameter Models, Hitting Memory Bottlenecks

pstAsiatech · x · 2026-08-12

As Qwen and Kimi teams push model sizes toward the 10-trillion (10T) parameter range, Alibaba's new generation SuperNode must scale to support them. Even at FP8 precision, running 10T models presents massive challenges.

Comparing the currently scaling Zhenwu M890 SuperNode with NVIDIA's GB300 NVL72, both systems will hit a physical memory capacity bottleneck before reaching their compute limits when supporting next-generation massive models.

Original post →

More from Infra

Infra channel →