Moore Threads Open-Sources Megatron-MUSA-Titan AI Stack
The new training software enables the development of massive trillion-parameter models on large GPU clusters.
Updated on Oct. 29, 2026 in Artificial Intelligence
Live Poll
Do you trust that domestic AI hardware can match the performance of international industry standards?
Moore Threads has released its Megatron-MUSA-Titan training stack as open-source software on GitHub. This framework is designed to facilitate the training of large-scale, trillion-parameter mixture-of-experts models.
Why it matters
The release provides specialized optimizations that allow developers to leverage Moore Threads hardware for massive AI workloads. By bridging the gap between MUSA architecture and existing large-language model frameworks, it lowers barriers for large-cluster model development.
The stack supports training on clusters exceeding 10,000 GPUs and targets models with up to 1 trillion parameters. It integrates blockwise FP8 training and utilizes the DeepEP engine for accelerated token management.
The players
Moore Threads
Moore Threads is a Chinese technology company specializing in the development of high-performance graphics processing units and artificial intelligence hardware.
GitHub
GitHub is a global web-based platform that provides hosting for software development and version control using Git.
The details
Built upon the NVIDIA Megatron-LM framework, the software uses the torch_musa PyTorch integration and MUSA C CUDA compatibility to manage high-demand training tasks. It also includes support for slime and miles reinforcement learning frameworks alongside a pretraining recipe for 1T MoE models.
Timeline
Moore Threads open-sourced the Megatron-MUSA-Titan training stack on October 29, 2026.
The Tech Race
This release marks a significant effort to establish a standardized software ecosystem around proprietary Chinese GPU hardware in a field long dominated by Western architectures. It replaces fragmented, manual workarounds with a unified framework to compete in the race to build trillion-parameter models.
For developers and researchers, this release provides a direct pathway to utilize high-end cluster hardware for training state-of-the-art AI models without proprietary restrictions. It simplifies the transition to Moore Threads hardware by allowing teams to port existing PyTorch-based workflows more efficiently.
The takeaway
Open-sourcing these tools lowers the barrier for academic and industry researchers to explore massive model architectures on diverse hardware. Organizations looking to scale their AI training operations can leverage these recipes to reduce the overhead of custom software integration.
Further reading
For more information on the latest developments in large-scale model training, visit Artificial Intelligence.
Live Poll
Do you trust that domestic AI hardware can match the performance of international industry standards?
