StepFun released Step 5 Preview open weights on Hugging Face on October 15, 2026, four hours behind its initially promised October 15 release date, underscoring the importance of treating vendor timelines as targets rather than guarantees. The model employs a Sparse Mixture-of-Experts (MoE) architecture with 600 billion total parameters but only 27 billion active parameters per token, enabling cost-efficient inference at $1.00 per million input tokens and $2.70 per million output tokens via API, or through OpenRouter. Three weight formats—FP8 (gated by manual approval), BF16, and NVFP4—are available under the StepFun Model License 1.0, which permits free use and modification with a required attribution notice for projects exceeding 100 million monthly active users or $20 million in monthly revenue. NVIDIA reported that the NVFP4 variant achieves nearly 8,000 tokens per second per GPU on GB300 hardware, though this figure represents a vendor-claimed peak under ideal batched conditions using 4-bit quantization, not independent verification or single-sequence performance. The model supports a 1-million-token context window and accepts text and image inputs via API, with video support added. Self-hosting requires substantial infrastructure, with a minimum recommended setup of four B300 GPUs in FP8 configuration. Benchmark results, including a 67.7 score on DeepSWE v1.1 and 33.3 on Terminal-Bench 4.0, were company-reported and not directly comparable to competitors due to differing reasoning effort levels. An independent Intelligence Index score of 43.7 from Artificial Analysis places Step 5 Preview among the top-performing open-weight models, validating its general capability without vendor bias.