Jieyue Xingchen (阶跃星辰) has officially released the Step 5 Preview model weights, marking a significant step forward in AI accessibility. The release includes two versions: BF16 and FP8, showcasing cutting-edge performance in AI programming, software engineering, professional knowledge work, and finance.
Open Source Release and Accessibility
The model weights are now available on Hugging Face and ModelScope. Developers can leverage vLLM or SGLang to deploy API services compatible with OpenAI interfaces, or use Transformers for offline inference. This move aims to democratize access to advanced AI capabilities and foster broader innovation within the developer community.
Industry-Wide Support and 0-Day Adaptations
In response to the release, leading technology companies including Huawei, Biren Technology, Muxi Systems, and Moore Threads have announced immediate 0-Day support for the Step 5 Preview model. This rapid adoption highlights the growing ecosystem of AI hardware and software vendors eager to integrate and optimize the latest advancements.
Huawei’s Ascend Platform Integration
Huawei has demonstrated a swift integration of the Step 5 Preview model onto its Ascend platform. Leveraging the vLLM inference engine, Ascend has achieved 0-Day support, with optimized inference tailored to the model’s architecture and the Ascend hardware’s unique characteristics.
Coinciding with the open-source release, Huawei activated its first 1024-card Atlas 950 SuperPoD liquid-cooled super-node on October 15th. This powerful system matches the performance of advanced overseas computing power, offering 1 EFLOPS FP8 and 2 EFLOPS FP4 precision, ideal for trillion-parameter large model inference.
The Ascend Atlas 950 SuperPoD features a 1024-card configuration, built on the Lingqu high-speed interconnect architecture. It boasts 256TB of global unified memory addressing, TB-level interconnect bandwidth between NPU cards, and latency as low as 3μs, effectively overcoming traditional communication bottlenecks in compute and storage resources.
Huawei explained that the Ascend 950 super-node, in conjunction with the vLLM Ascend inference engine and the Step 5 Preview model, has undergone extensive optimization. This includes integrating Ascend C and Triton fused operators, optimizing multi-stream parallelism and compute bubbles, and implementing speculative decoding and NPU Graph execution. This holistic approach ensures efficient computational support for the Step 5 Preview deployment.
Muxi Systems’ Commitment to the Ecosystem
Muxi Systems has also confirmed Day 0 deep adaptation for the Step 5 Preview model on its C-series GPUs. With its fully self-developed computing power foundation, Muxi aims to bolster the domestic Agent ecosystem and maintain its leadership in rapid adaptation of Chinese large models.
Since December 2025, Muxi has reportedly completed Day 0 adaptations for 43 mainstream flagship models from leading vendors like Zhipu AI, Alibaba’s Qwen, MiniMax, DeepSeek, Jieyue Xingchen, and Tencent’s Hunyuan. This rapid and extensive adaptation pace positions Muxi at the forefront of the industry.
The Muxi C-series GPUs are built on the self-developed XCORE architecture, featuring high-bandwidth memory and MetaXLink high-speed interconnect. They offer hybrid-precision computing power and linear scalability, naturally supporting the demands of MoE model expert scheduling and million-token context length inference, whether in single-machine multi-card or large-scale intelligent computing clusters.
On the software front, Muxi’s self-developed MXMACA full-stack software suite, encompassing drivers, the MXCC compiler, optimized operator libraries, and AI framework adaptation layers, ensures native compatibility with mainstream AI development ecosystems. This significantly reduces model migration costs, compressing adaptation cycles from weeks to hours.
Biren Technology’s Accelerated Inference
Biren Technology has completed the Day 0 adaptation for the Step 5 Preview model on its Biren series GPUs. Utilizing the vLLM inference framework and its mature BIRENSUPA software stack, Biren has enabled efficient text understanding, generation, and inference, validating key processes from weight loading to service startup and text generation.
Furthermore, Biren’s self-developed SUPACODE platform, a full-stack multi-agent system, can automatically parse and adapt critical modules of Step 5 Preview, such as DSA, SummaryCache, and MoE, to streamline the entire inference pipeline.
Moore Threads’ Day-0 Adaptation
Moore Threads has also achieved Day-0 rapid adaptation and efficient operation of the Step 5 Preview model on its flagship MTT S5000 AI training and inference card, supported by the MUSA software stack. The adaptation and optimization cover hardware computing power, the software stack, and the inference framework, precisely tailored to the model’s core capabilities.
The open-sourcing of Jieyue Xingchen’s Step 5 Preview model weights, coupled with the immediate and robust support from major hardware vendors, signifies a vibrant and rapidly evolving AI landscape. This collaboration is expected to accelerate the development and deployment of advanced AI applications across various industries.
