StepFun opens Step 5 Preview model weights, Chinese chip makers offer day-one support

StepFun has officially released the weights of its Step 5 Preview model in BF16 and FP8 versions, now available on Hugging Face and ModelScope. The model performs at the frontier in AI coding, software engineering, and finance tasks, supporting a 1 million token context window.

Huawei's Ascend chips offered 0-day support via the vLLM inference engine. The first 1024-card Atlas 950 SuperPoD liquid-cooled super node also went live, providing 1 EFLOPS of FP8 computing power and 256TB of unified memory addressing. Biren, MetaX, and Moore Threads also completed Day 0 adaptation on their own GPUs.

MetaX said it has completed Day 0 adaptation for 43 mainstream flagship models since December 2025, covering Zhipu, Alibaba Qwen, DeepSeek and more. This synergy between open models and rapid chip adaptation provides integrated software and hardware support for China's AI ecosystem.