Several domestic computing platforms moved quickly after StepFun opened the weights of Step 5 Preview on October 15. Huawei, MetaX, Biren and Moore Threads each announced Day-0 support for the model the same day, meaning it could be loaded and served on their hardware from release.
Huawei said its Ascend platform supported the model through the vLLM Ascend inference engine and had built an optimization package around the model's architecture and the Atlas 950 SuperPoD hardware. The work spans fused operators written in Ascend C and Triton, multi-stream parallelism, reduced compute bubbles, MTP speculative decoding and NPU graph execution.
Huawei also said the first 1,024-card Atlas 950 SuperPoD liquid-cooled system was lit up on October 15 alongside the Step 5 Preview release. The SuperPoD product line was announced earlier; what is new is that the first full 1,024-card unit is now running, with Step 5 Preview as its launch workload. Huawei describes the system as delivering 1 EFLOPS at FP8 and 2 EFLOPS at FP4, with 256TB of globally unified memory addressing and interconnect latency as low as 3 microseconds, and says its performance is on par with advanced computing systems elsewhere. Those figures are Huawei's own.
MetaX said its C-series GPUs completed Day-0 adaptation using its in-house MXMACA software stack. The company says it has delivered Day-0 adaptation for 43 flagship models since December 2025, including models from Zhipu, Alibaba's Qwen, MiniMax, DeepSeek, StepFun and Tencent Hunyuan.
Biren said it adapted the model on its general-purpose GPUs using vLLM and its BIRENSUPA software stack, bringing up weight loading, service startup and text generation and completing function and accuracy checks. The company said its SUPACODE multi-agent platform automatically parsed and adapted key modules of the model, including DSA, SummaryCache and the MoE layers.
Moore Threads said its MTT S5000 training and inference card and MUSA software stack ran the model on day one, with tuning across hardware, software and the inference framework.
Same-day support from multiple vendors is becoming a routine part of major open-weight releases in China, giving developers more choice of hardware for deployment from launch. For details of the model itself, see our coverage of the open-weights release.
SEE ALSO: StepFun Releases Open Weights for Step 5 Preview, Publishes Technical Blog on Its Architecture · Huawei to Publicly Demonstrate Atlas 950 SuperPoD AI Computing Hardware at WAIC 2026