The premier large-scale model, Step 5 Preview, unveiled by Step-General AI, employs a sparse Mixture of Experts (MoE) framework. This model boasts a staggering 600 billion parameters in total, yet only 27 billion are activated during each operation. It accommodates a context length of up to 1 million Tokens and supports both image and text input. In terms of performance metrics, Step 5 Preview stands toe-to-toe with K3, which boasts a massive 2.8 trillion parameters, despite having a significantly smaller parameter count of 600 billion. It achieved a commendable score of 44 on the AA Benchmark Intelligence Index, matching the performance of Kimi K3 Max. Its output speed is impressive, clocking in at approximately 100 Tokens per second, with a cost-effective rate of $0.71 per task. The model excels in handling long-duration tasks, can autonomously and continuously optimize GPU kernels for 24 hours, and is capable of independently designing training data to boost the accuracy of other models. Currently, both the API and Studio services for Step 5 Preview are fully accessible, with the full-weight model set to be released on October 15.