Overview
- Jieyue Xingchen published Step 5 Preview model weights and technical materials under open terms, offering BF16 and FP8 versions plus documentation and developer tools for immediate use.
- The model uses a sparse Mixture‑of‑Experts design with about 600 billion parameters, a 27B per‑token activation footprint, and native text‑and‑vision inputs to support million‑token context windows for long‑range agent tasks.
- Weights are available on public hubs and the model is compatible with common runtimes such as vLLM, SGLang, and Transformers, allowing deployment as an OpenAI‑compatible API or for offline inference.
- Multiple domestic vendors announced Day‑0 adaptations and optimizations—most notably Huawei lighting a 1024‑card Atlas 950 SuperPoD node with large unified memory and EFLOPS‑class FP8/FP4 claims—to enable practical inference at scale.
- While the open release lowers software barriers for researchers and developers to build long‑context agents, real‑world use will depend on access to specialized hardware and continued optimization because million‑token MoE workloads remain very resource‑intensive.