StepFun Releases Weights for Flagship Model Step-5-Preview

StepFun has released the weights of its flagship sparse MoE model Step-5-Preview — 600 billion parameters with 27 billion active per token, 1 million token context, and multimodal input.

Compiled by Sergey KostenchukPublished 2026-09-24Updated 2026-09-24

2026-09-24 Research OpenAI

🤖 StepFun Releases Weights for Flagship Step-5-Preview

The weights are available on Hugging Face (checkpoint TypeSafeAI/Step-5-Preview-BF16, released September 20, 2026). This is a sparse MoE: 600 billion parameters, with only 27 billion active per token (~4.5%). Context — 1 million tokens, input — text, images, and videos up to 128 MB.

🌍 StepFun jumped from Step-3.7-Flash to Step 5, skipping Step 4.x, with the 'Pareto Frontier' philosophy. An Intelligence Index of 44 is claimed — top 3 among open-weight. Inference — 8× H100 in BF16. This puts price pressure on Claude Opus 5 and GPT-6 Astra and changes the economics of the segment where Kimi K3 Max carries ~2.8T parameters.

👤 The weights can be run locally via vLLM or SGLang, or an OpenAI-compatible API api.stepfun.com (step-5-preview) is available: parallel tool calls, strict JSON Schema, reasoning_effort levels. There are no independent measurements of near-frontier quality at such sparsity yet.

Source 1: https://huggingface.co/TypeSafeAI/Step-5-Preview-BF16

Sources