StepFun has recently officially released and open-sourced the weights of the Step5Preview model, simultaneously making the BF16 and FP8 model weights, a detailed technical blog, and accompanying open-source development tools available to developers worldwide. Previously, thanks to its outstanding overall performance, the model once ranked second globally in call volume on OpenRouter's weekly and daily leaderboards.
In terms of architectural design, Step5Preview is deeply optimized for long-horizon complex Agent tasks, adopting an advanced sparse MoE (Mixture of Experts) architecture. The model has a total parameter count of up to 600B, but only 27B parameters are activated per token, ensuring powerful reasoning capability while achieving extremely high operational efficiency. It also natively supports text and vision multimodal input, and features an astonishing 1 million token ultra-long context window, allowing it to handle a wide range of professional and complex scenarios with ease.
To make it easy for developers to get started quickly and integrate into their ecosystems, the model's weights are now fully available on the Hugging Face and ModelScope platforms, perfectly compatible with mainstream chip platforms and inference frameworks, and supporting diverse local deployment and offline inference methods, greatly facilitating developers in integrating it into their own systems to build various advanced intelligent applications.
At the same time, the technical blog published by StepFun details its key innovations in joint optimization of architecture design and training, revealing the technical foundation that supports the stable operation of long-horizon complex Agents. With the weights now fully open, this high-performance model is expected to further accelerate the deployment and innovation of domestic large models in various vertical and long-horizon application scenarios.