Moore Threads Launches Open-Source Megatron-MUSA-Titan Training Infrastructure

Moore Threads has announced the launch of the open-source Megatron-MUSA-Titan infrastructure, designed for training large models on over 10,000 GPUs. The project targets trillion-parameter mixture-of-experts models and covers both pretraining and reinforcement learning. The infrastructure utilizes NVIDIA's Megatron-LM framework, allowing developers to run training on Moore Threads GPUs without modifying upstream code. The repository includes pretraining recipes for a Kimi-K2-style 1T MoE model as an example configuration.