Moore Threads Open-Sources Megatron-MUSA-Titan for Trillion-Parameter MoE Training on Its GPUs

Moore Threads has open-sourced Megatron-MUSA-Titan, the training stack it uses for trillion-parameter MoE pretraining on 10,000-GPU clusters, with blockwise FP8 and support for slime and miles RL.

NewsBeagle is a discovery engine. We show the headline, a short source-supplied summary and clear attribution. The complete reporting remains with the publisher.