StepFun opens Step 5 Preview weights: 600B sparse MoE (27B active), 1M context, FP8/BF16 on Hugging Face and ModelScope

Why it matters

【Key points】

StepFun opened the weights for its flagship Step 5 Preview on 2026-10-15 as promised. Hugging Face repos stepfun-ai/Step-5-Preview (FP8) and stepfun-ai/Step-5-Preview-BF16 are ungated with downloadable shards; ModelScope is live too. Per the model card: sparse MoE, ~600B total / ~27B active per token, up to 1M context, text+image input, aimed at software engineering and long-horizon agents, with a finance focus. License is StepFun Model License 1.0 (Apache-2.0-style with attribution and commercial conditions). Serve via vLLM / SGLang or Transformers. The NVFP4 repo (created 10/14) is a quant side-car, not the headline.

【Verifiable】

HF FP8: https://huggingface.co/stepfun-ai/Step-5-Preview (lastModified 2026-10-15T01:28:10Z, gated=false)

HF BF16: https://huggingface.co/stepfun-ai/Step-5-Preview-BF16

Model page: https://www.stepfun.com/step-5-preview

Docs: https://platform.stepfun.ai/docs/en/guides/models/step-5-preview

ModelScope post: https://x.com/ModelScope2022/status/2110568054623031787 (~03:06Z)

Site previously had API/free-week distribution (e/657) and STEPX Neo phone (e/842), not open weights.

【Our read】

1. Hard open-weight delivery from a Chinese lab: promised for Oct 15, gate opened the same day, files and license checkable — home_week material.

2. Nemotron 3.5 Bolt still has no official HF repo today; community “one delivered, one missing” chatter is context only, not an NVIDIA release.

3. License is not pure Apache — commercial thresholds in LICENSE.txt matter for redistribution and products.

4. @StepFun_ai posted nothing this window; primary sources are HF/ModelScope and the model card.