- Moore Threads completed Day-0 adaptation of its MTT S5000 GPU and MUSA stack for StepFun's Step 5 Preview on the same day the model was open-sourced.
- The adaptation covers mixed attention, mixed precision inference and speculative decoding for the 600-billion-parameter sparse MoE model.
- Step 5 Preview natively supports an ultra-long context window of up to one million tokens.
Day-0 Hardware Adaptation
Moore Threads announced on October 15 that its MTT S5000 AI computing card and accompanying MUSA software stack had completed Day-0 adaptation for StepFun's Step 5 Preview model, which was open-sourced on the same day. The company said the adaptation work covered three core technical areas: mixed attention mechanisms, mixed precision inference and speculative decoding.
Step 5 Preview uses a sparse mixture-of-experts architecture with 600 billion total parameters and 27 billion active parameters per token, and natively supports a context window of up to one million tokens. Moore Threads framed the rapid adaptation as evidence of its ability to respond quickly when frontier domestic models are released.
Domestic Ecosystem Significance
The adaptation represents a maturing domestic AI hardware-software ecosystem in China. Moore Threads is positioning the MTT S5000 as a viable platform for deploying complex long-context models, offering developers an alternative to imported accelerators for running frontier Chinese LLMs.
The company described the collaboration as marking a new level in domestic intelligent-computing hardware's ability to support frontier large-model ecosystems, providing developers with a mature and reliable foundation for deploying and running complex long-text models on domestic compute platforms.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01
