Moore Threads MTT S5000 Adapts FlagOS Training Stack, Enables Uninterrupted Qwen3-0.6B Training

The Beijing Academy of Artificial Intelligence recently released a verification result: Zhongzhi FlagOS completed the "full-element" verification of AI training with a unified technology stack. As a core participant, Moore Threads fully adapted the FlagOS training full-element software stack based on its MTT S5000 GPU.

In this verification, MTT S5000 completed the 1T Tokens training of Qwen3-0.6B model from scratch, running stably for over 14000 steps without interruption for 6 consecutive days. The training Loss curve is highly consistent with the baseline, and the downstream task evaluation is 1.65 percentage points higher than the industry benchmark, verifying its computing power stability and effectiveness.