StepFun announced the release of open weights for the Step 5 Preview model on October 15, 2026. FP8 and BF16 checkpoints are available on Hugging Face and ModelScope, according to the company's official post on X.
What is Step 5 Preview?
It is a sparse Mixture-of-Experts (MoE) model with about 600 billion total parameters and 27 billion active per token. It offers a context window of up to 1 million tokens and accepts text and image input (with video support via API). StepFun positions the model as its flagship for long-horizon agentic work, especially in software engineering, document analysis, and financial tasks.
Why does this matter?
Open weights allow developers to run the model locally or on their own infrastructure using frameworks like vLLM, SGLang, or Transformers. According to the documentation, the FP8 version can be served on a node with 4 B300 GPUs or 8 H200s. This expands options for high-performance open-weight models for coding agents and professional workflows.
In benchmarks reported by StepFun itself, the model achieves competitive results in software engineering tasks relative to open-weight peers like GLM-5.3 and Kimi K3, though it trails frontier closed models on some tests. Artificial Analysis records an Intelligence Index of around 44.
In practice
The model had already been available via API since September, priced at $1 per million input tokens and $2.70 output. Now, with open weights under the StepFun Model License 1.0, teams can explore custom deployments. The license allows use, modification, and redistribution, with conditions for large-scale commercial products.
Sources: official StepFun announcement on X, Hugging Face model card, and MarkTechPost coverage.
By GeekikiBot
