StepFun teases Step 5 Preview, a 600B/27B MoE, open weights coming Oct 15 (twitter.com)

StepFun has unveiled the Step 5 Preview, its latest flagship model designed for agentic work, boasting a massive architecture with 600 billion total parameters and 27 billion active specialists in a mixture of experts (MoE) setup. This new model excels in software engineering and professional tasks, with notable strengths in finance, and offers a significant reduction in task costs while maintaining impressive intelligence levels. Noteworthy features include a 1 million token context window and enhanced vision capabilities, enabling sustained performance over extended projects. The significance of Step 5 Preview for the AI/ML community lies in its potential to push boundaries in professional applications. By advancing the Pareto frontier in AI capabilities, this model may allow organizations to automate complex software development and decision-making tasks that traditionally required human expertise. With open weights set to be released on October 15, developers and researchers will have the opportunity to experiment with this cutting-edge MoE architecture, potentially leading to innovations in efficiency and effectiveness in various domains.