Step 5 Preview, a 1M-context MoE from StepFun, shows up on OpenRouter (openrouter.ai)

StepFun has launched Step 5 Preview, a groundbreaking model designed for agentic work that boasts a sparse Mixture-of-Experts (MoE) architecture with 27 billion active parameters out of a total of 600 billion. This model excels in complex software engineering tasks and professional environments, particularly in the finance sector, where it shows distinctive capabilities in managing extensive codebases and documents. It is built to facilitate multi-step processes, refining results through integrated tools, which is significant for developers seeking advanced AI assistance in intricate projects. Additionally, StepFun introduced Step 3.7 Flash, an advanced multimodal MoE model that fuses a 196 billion-parameter language backbone with a vision encoder for enhanced understanding of images and videos. This model features a 256,000-token context window and various reasoning levels, allowing users to balance speed, cost, and depth of reasoning efficiently. The focus on high-efficiency and selective parameter activation across these models highlights a trend in AI to optimize resource usage while enhancing functional capabilities, paving the way for more powerful and versatile AI applications in both coding and structured output tasks.