On November 17, Moore Threads announced that its laboratory had completed adaptation of the training pipeline for Moonshot AI's Kimi K3.1 Pro model, running 400 consecutive steps on a two-server, 16-card configuration without numerical overflow or abnormal interruption.
Kimi K3.1 Pro is a 2.8-trillion-parameter mixture-of-experts native multimodal model whose dynamic sparse attention and expert routing place high demands on multi-card communication and memory management. The test used a trimmed configuration - 13 language layers, 16 routing experts with top-eight activation, and a complete 27-layer vision tower - trained from random initialisation on a visual question-answering dataset. Training loss fell from 8.9365 to 2.1802, with peak memory use of 44.80 GiB.
The company stressed the work is a pipeline verification under a trimmed configuration, not a full-scale pre-training run. The announcement comes as domestic GPUs accelerate adaptation to leading Chinese models. TrendForce forecasts China will ship about five million high-end AI chips in 2026, up 83 percent, with domestic solutions nearing a 90 percent share.





