On October 15, 2026, StepFun made a significant announcement—the official release of the Step 5 Preview weights. These weights come in both BF16 and FP8 precision formats, catering to a wide range of computational needs. Accompanying this release is a comprehensive technical blog, along with a suite of open-source development tools and deployment guides, designed to facilitate seamless integration and utilization.

The Step 5 Preview model boasts a sophisticated sparse Mixture-of-Experts (MoE) architecture, housing a staggering 600 billion parameters. However, it cleverly activates only 27 billion parameters per token, optimizing computational efficiency without compromising on performance. This model is engineered to support an extensive context window of up to 1 million tokens, enabling it to process and understand vast amounts of information with ease. Moreover, it natively handles both textual and visual inputs, making it a versatile tool for a multitude of applications, including software engineering, professional knowledge work, and finance.

The model weights are readily accessible on popular platforms such as Hugging Face and ModelScope, ensuring widespread availability and ease of access. Furthermore, they have been meticulously adapted to support multiple chip platforms and inference frameworks, enhancing their compatibility and flexibility.

According to the latest data from OpenRouter as of October 15, the Step 5 Preview has made a remarkable impact, securing the second position globally on the platform's weekly and daily Token usage charts. This achievement underscores its growing popularity and the trust that users place in its capabilities.