StepFun introduced the Step 5 Preview model with 600 billion parameters

Chinese company StepFun announced the release of a new multimodal model Step 5 Preview. The system is built on the Mixture-of-Experts architecture and contains 600 billion parameters, providing support for a contextual window of up to 1 million tokens.

Developers focused on specializing the model in software engineering and financial analysis. According to internal tests, Step 5 Preview demonstrates an advantage over current competitors in specialized benchmarks, including DeepSWE, ProgramBench, and DRACO. High performance in coding tasks and processing complex financial data is the key development vector for this version.

The model's technical characteristics allow it to efficiently process long sequences of data, which is critical for analyzing large code repositories and voluminous reports. The integration of the Mixture-of-Experts architecture provides a balance between computational power and response generation speed when working with complex queries.

The release of Step 5 Preview is aimed at strengthening the company's position in the corporate AI solutions segment. Detailed discussions of technical aspects and initial testing results by the community are available on the Hacker News platform. The implementation of such systems allows for the automation of complex software development stages, reducing debugging and documentation writing time.