Step 5 Preview

600B total / 27B active model with downloadable FP8/BF16 weights, conditional StepFun licensing and large-memory serving.

Model Specificationsstepfun

FP8/BF16 multi-GPU configurations; weight memory and KV cache are separate.

Self-hosting costs depend on hardware and runtime; no hosted rate is inferred from weights.

Scoped Terms

🚀Function Overview

Primary-source contract checked October 21, 2026. No independent model request or hardware test was executed.

Official repositories: stepfun-ai/Step-5-Preview and stepfun-ai/Step-5-Preview-BF16.
Up to 1M combined context; text/image input and text output in the specification table.
The vLLM/SGLang local recipes use served-model-name step5 and OpenAI-compatible requests; this is not evidence of a public hosted alias.
FP8: 4 × B300 or 8 × H200; BF16: 8 × B300 or 16 × H200 across two nodes.
The card says the BF16 standalone conversion-to-generation path has not been verified end to end.
The September 20 introduction remains historical; the October 15 official repository update and actual weight shards support a separate weight milestone.

string

stepfun-ai/Step-5-Preview-BF16 — follow the selected official runtime or provider interface; this is not a universal hosted alias.

provider-specific

Use the documented input schema; see official sources for supported modalities and request fields.

Hardware Type: FP8/BF16 multi-GPU configurations; weight memory and KV cache are separate.
Commercial Use: StepFun Model License 1.0: notices, attribution and conditional commercial terms, including hardware makers and qualifying MaaS businesses; see the practical guide.
Pricing: Self-hosting costs depend on hardware and runtime; no hosted rate is inferred from weights.