MUSA

Headlines · 3

1. 4d
   Moore Threads Open-Sources Megatron-MUSA-Titan for Large-Scale MoE Training

   Moore Threads announced the open-sourcing of Megatron-MUSA-Titan, a large-model training infrastructure project adapted for the company’s MUSA platform and full-function GPUs. IT之家 reported that it targets trillion-parameter MoE pretraining and reinforcement learning on clusters with 10,000 GPUs. The project supports blockwise FP8 training and can work with the DeepEP and ACE asynchronous communication engines to overlap expert parallelism and computation. Moore Threads said the project had been used in continued trillion-parameter MoE pretraining on its 10,000-GPU-scale cluster and supports coordination with slime and miles, Megatron training, and SGLang sample generation. Its reported model FLOPS utilization and long-running stability were not independently verified in the supplied reports.

   IT之家 AI · 🔥 0
2. 4d
   Moore Threads posts first quarterly non-GAAP profit on surging AI GPU revenue

   Moore Threads reported third-quarter revenue of RMB 1.552 billion, up 1,773.92% year over year, with attributable net profit of RMB 49.26 million and non-GAAP net profit of RMB 20.07 million. Revenue for the first nine months reached RMB 3.289 billion, up 319.15%, while attributable net profit was RMB 37.70 million. However, the nine-month non-GAAP result remained a RMB 131 million loss, gross margin fell to 53.71%, and operating cash flow was negative RMB 3.309 billion, according to 36Kr, TMTPost and IT之家 AI.

   钛媒体 · 🔥 0
3. Oct 16
   Moore Threads says MTT S5000 gains Day-0 support for StepFun's Step 5 Preview

   AIbase AI新闻 reported on Oct 16 that StepFun open-sourced its new Step5Preview model on Oct 15, and that Moore Threads announced the same day that its MTT S5000 compute card and MUSA software stack had completed Day-0 adaptation for the model. The report says Step5Preview uses a sparse MoE architecture with 600B total parameters and 27B activated per token, and natively supports context windows up to 1 million tokens; Moore Threads said the work covered hybrid attention, mixed-precision inference and speculative decoding. Both the depth of the adaptation and the claim that domestic compute now supports frontier-scale models come from the vendor, with no third-party benchmarks or independent verification in the report.

   AIbase AI新闻 · 🔥 0

Experience and discussion from the community

Share my MUSA experienceAsk about MUSA