stepfun-ai/Step-5-Preview model card

Fetched 2026-10-18. Weights repo created 2026-09-20 (HF API createdAt 2026-09-20T03:15:06Z). Card nameplate: sparse Mixture-of-Experts, 600B total parameters, 27B active per token, context window up to 1M tokens, text and image in, text out. The card's numeric benchmark table is an image, not text; the numbers cited from it are transcribed from the vendor model page, not read off this card. Deployment section lists vLLM and SGLang serving recipes and a Transformers offline path. The smallest published configuration is FP8 on one B300 node with 4 GPUs (TP4, TEP4 or DEP4), or FP8 on one H200 node with 8 GPUs. BF16 is listed at 8 B300 GPUs or 16 H200 GPUs. No llama.cpp, GGUF, or single-GPU expert-offload recipe is on the card. License files LICENSE.txt and LICENSE-CN.txt are in the repo.