StepFun Open-Sources Step 5 Preview Model with 0Day Domestic Hardware Support
StepFun has officially open-sourced the Step 5 Preview model weights in BF16 and FP8 versions, receiving immediate Day-0 support and performance optimization from domestic Chinese hardware vendors including Huawei Ascend, MetaX, Biren Technology, and Moore Threads. This development marks a major milestone in integrating domestic AI hardware ecosystems with frontier large language models, demonstrating that domestic accelerators can achieve rapid day-one adoption and competitive performance for complex AI workloads. The model weights are available on Hugging Face and ModelScope, supporting deployment via vLLM or SGLang with OpenAI-compatible APIs. Concurrently, Huawei lit up its first 1024-card Atlas 950 SuperPoD liquid-cooled supernode cluster to provide robust computing support.
## BACKGROUND
StepFun is an artificial intelligence startup focusing on developing advanced large-scale foundational models and agentic AI systems. Hardware vendors like Huawei Ascend build specialized AI processors and supernode architectures, such as the Atlas series, designed to handle massive computational demands of modern trillion-parameter models.