Moore threads launches open-source Megatron-MUSA-Titan for training trillion-parameter moe models

On October 28, 2026, Moore Threads announced the open-sourcing of its Megatron-MUSA-Titan project, a large model training infrastructure designed for trillion-parameter MoE models on multi-thousand GPU clusters. The project supports both pre-training and reinforcement learning scenarios, offering developers training adaptation and system optimization capabilities for the MUSA platform. Megatron-MUSA-Titan has been continuously applied in pre-training trillion-parameter MoE models on Moore Threads' multi-thousand GPU clusters, achieving high MFU and long-term stable training. It also supports integration with slime and miles reinforcement learning frameworks, combining Megatron training with SGLang sample generation to create a training foundation spanning pre-training and reinforcement learning. The project supports blockwise FP8 training, leveraging Moore Threads' low-precision computing capabilities to improve computational efficiency in large-scale training. It also integrates with DeepEP and ACE asynchronous communication engines to accelerate token distribution and result aggregation in MoE models, enabling expert parallelism and overlapping computation with data exchange to enhance end-to-end training efficiency.