Moore Threads Releases Megatron-MUSA-Titan Training Stack
The new open-source framework enables large-scale training of trillion-parameter models on Moore Threads hardware 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 open-sourced its Megatron-MUSA-Titan training stack, a software architecture designed to facilitate the training of trillion-parameter mixture-of-experts (MoE) models. The repository, now available on GitHub, supports scaling to clusters exceeding 10,000 GPUs.
Why it matters
The stack provides critical optimizations for Moore Threads hardware, addressing MUSA-specific compatibility issues to enable high-performance model training at scale. By adapting the Megatron-LM framework, it expands the software ecosystem available for developers working with specialized hardware clusters.
The framework enables training for 1-trillion parameter MoE models across clusters exceeding 10,000 GPUs. Performance is supported by blockwise FP8 precision through the MT-TransformerEngine.
The players
Moore Threads
A developer of specialized GPU hardware and software stacks aimed at large-scale artificial intelligence training.
The details
Megatron-MUSA-Titan adapts the NVIDIA Megatron-LM framework to run on Moore Threads hardware by utilizing torch_musa PyTorch integration and MUSA C, a layer providing CUDA compatibility. It optimizes Mixture-of-Experts (MoE) token dispatch and combine operations using DeepEP, which executes over the hardware's Async Copy Engine to reduce latency. The software also integrates support for the slime and miles reinforcement learning frameworks.
Timeline
October 29, 2026: Moore Threads released the source code on GitHub.
The Tech Race
The release follows the industry pattern of extending the NVIDIA Megatron-LM framework to support diverse, non-NVIDIA hardware clusters. It marks a push to standardize large-model training workflows across specialized silicon architectures.
Developers and AI researchers can now access the pretraining recipe for 1-trillion parameter MoE models directly via the GitHub repository. The software requires a MUSA-compatible environment to leverage the specific cluster-scale optimizations provided.
The takeaway
This release accelerates the accessibility of large-scale MoE training for teams utilizing Moore Threads hardware. Future benchmarks comparing this implementation against established H100 or B200 training clusters will be the next milestone to verify performance parity.
Further reading
Explore more developments in the Artificial Intelligence sector.
Source note: This article includes information reported by Pandaily.
Live Poll
Do you trust that domestic AI hardware can match the performance of international industry standards?
