StepFun opens Step 5 Preview's weights: 600B total, 27B active, 1M context, under its own license

StepFun released the weights of Step 5 Preview on October 15, the date it set when the model went on sale through its API on September 20. FP8 and BF16 checkpoints are on Hugging Face and ModelScope, and an NVFP4 build is on Hugging Face too.

THE MODEL

A sparse mixture-of-experts with 600B parameters in total and 27B active per token; the safetensors index on Hugging Face counts 604.3B. The model card lists up to 1M tokens of context and up to 1M of output, with text and image input. StepFun pitches it at long-horizon agentic work: repository-level coding, professional knowledge work and, in particular, finance. Through StepFun's API it costs $1.00 per 1M input tokens and $2.70 per 1M output.

WHAT IT TAKES TO RUN

The FP8 files come to about 650 GB and the BF16 files to about 1.2 TB. StepFun's deployment table lists one 8-GPU H200 node or four B300 GPUs for FP8 with vLLM or SGLang, and two H200 nodes for BF16. It is a data-center model, not one for a desktop.

THE LICENSE

It is not Apache 2.0 or MIT. The StepFun Model License 1.0 lets anyone use, modify and redistribute the weights for free, with two conditions on top. A commercial product with more than 100M monthly active users or more than $20M in monthly revenue must show "Powered by StepFun AI". And two kinds of company need a separate commercial license from StepFun before they ship: device makers putting the model in phones, cars, glasses or robots, and model-as-a-service companies with more than $20M a year in revenue. Internal use is exempt from the second condition. The license also treats a model trained to a material extent on Step 5's outputs as a derivative work.

STEPFUN'S OWN NUMBERS

On StepFun's chart, Step 5 Preview on its High setting beats Kimi K3 and GLM-5.3 on most rows: DeepSWE v1.1 67.7 against 67.5 and 66.9, ProgramBench 80.5 against 77.8 and 72.0. GLM-5.3 beats it on Terminal-Bench v4 (41.9 against 33.3) and GDPval-AA v2.1 (1645 against 1566). The closed models on the chart are Claude Opus 5 and GPT-6 Astra, not Opus 5.5, and they lead on most rows, including DeepSWE (74.0 and 74.1) and Terminal-Bench v4 (52.3 and 57.9). These are StepFun's figures, not independent results.

WHAT WE DO NOT KNOW

Whether a final Step 5 follows the Preview, and whether it will be open too. Step 5 Preview is not yet in Epoch AI's Capabilities Index, and no independent lab has published results on the open checkpoint.