Moore Threads open-sources Megatron-MUSA-Titan for trillion-parameter AI training

Moore Threads has announced the open-source release of Megatron-MUSA-Titan, a training infrastructure project for large AI models. Built on Moore Threads' full-function GPUs, it targets trillion-parameter MoE model training on 10,000-GPU clusters, covering both pre-training and reinforcement learning, with training adaptation and system optimization for the MUSA platform. The project has been used in long-running pre-training of trillion-parameter MoE models on Moore Threads' large clusters, achieving high computing utilization and stable training. It also works with reinforcement learning frameworks like slime and miles, connecting pre-training and reinforcement learning in one pipeline. On the technical side, it supports blockwise FP8 training to bring low-precision computing into the process. It can also pair with DeepEP and the ACE communication engine to speed up token distribution and result merging in MoE models, improving end-to-end training efficiency.