Reportedly: StepFun open-sources Step 5 Preview weights; Huawei Ascend, MetaX, Biren provide day-zero support
Reported by IT Home · not yet confirmed by the company or a second independent outlet. We update this page when it is.
StepFun released Step 5 Preview model weights in BF16 and FP8 on Hugging Face and ModelScope. Huawei Ascend, MetaX, and Biren Technology announced day-zero support, with Huawei lighting up a 1024-card Atlas 950 SuperPoD liquid-cooled supernode, IT Home reported.
China context
- Original name
- 阶跃星辰
- Outside China
- Open weights · ithome.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download Step 5 Preview weights from Hugging Face and deploy via vLLM or SGLang with OpenAI-compatible APIs, or use Transformers for offline inference.
- For investors
- The simultaneous day-zero support from three Chinese chip vendors indicates a coordinated ecosystem play, potentially increasing the attractiveness of StepFun's models for enterprises in China and reducing reliance on NVIDIA.
Translated from Chinese. Quotes and facts link to the original sources.
StepFun announced the open release of Step 5 Preview weights in BF16 and FP8, targeting AI coding, software engineering, professional knowledge work, and finance. The weights are available on Hugging Face and ModelScope, deployable via vLLM or SGLang with OpenAI-compatible APIs, or offline inference with Transformers. Huawei, Biren Technology, and MetaX announced day-zero support. Huawei's Ascend implemented day-zero support based on the vLLM inference engine with optimizations for the model architecture and hardware. On October 15, the first 1024-card Atlas 950 SuperPoD liquid-cooled supernode went online simultaneously with the open-source release, with performance on par with overseas advanced compute. The Atlas 950 SuperPoD is based on the LingQu high-speed interconnect architecture and supernode technology, the industry's largest 1024-card supernode product, providing 1 EFLOPS FP8 and 2 EFLOPS FP4 compute, suitable for trillion-parameter model inference. It features 256TB global unified memory addressing and TB-level interconnect bandwidth between NPUs.
Step 5 Preview is released in BF16 and FP8 formats, enabling efficient deployment on both standard GPU stacks and domestic Chinese accelerators. Huawei's day-zero support via vLLM indicates pre-release collaboration or rapid adaptation, with optimizations for the model architecture and hardware. The Atlas 950 SuperPoD's 1024-card configuration with 1 EFLOPS FP8 and 2 EFLOPS FP4 compute targets large-scale inference for trillion-parameter models, suggesting Step 5 Preview may be a very large model. The 256TB unified memory and TB-level NPU interconnect bandwidth are designed to handle memory-intensive inference workloads.
Chinese AI chip vendors gain immediate compatibility with a frontier open-weights model, reducing the software barrier for enterprises that must use domestic accelerators. StepFun's open release expands the set of high-performance open models available to developers outside China, while Huawei's simultaneous supernode launch demonstrates that domestic hardware can serve large-scale inference for such models.
Enterprises can deploy Step 5 Preview on both NVIDIA and Chinese domestic hardware, potentially lowering inference costs and avoiding vendor lock-in. The day-zero support from Huawei, MetaX, and Biren suggests that StepFun may have coordinated with hardware partners, which could accelerate enterprise adoption in China. For developers outside China, the open weights provide an alternative to proprietary models for coding and professional tasks.
Watch for independent benchmarks of Step 5 Preview against other open models, and for adoption metrics on Hugging Face and ModelScope. A key signal will be whether other Chinese chip vendors (e.g., Biren, MetaX) publish detailed performance benchmarks or deployment guides. Also monitor whether StepFun releases larger or instruction-tuned variants.