According to CoinMeta, Step Star has released its flagship models Step and preview, which adopt a sparse moe architecture with a total of 600b parameters. Each activation consumes 27b of resources and supports 1 million token contexts as well as text and image inputs. Models API and Studio are now fully available, and the complete model weights will be released on October 15th. Early evaluations of Artificial Analysis have been conducted; Step scored 44 points in Intelligence Index, tying with Kimi K3 Max, and produced approximately 100 token per task, with a cost of 0.71 US dollars per task, while Kimi K3 Max cost 2 US dollars. The official demonstration showcased its capability for long-duration tasks. Step continuously optimized its H100 GPU kernel for 24 hours, reaching 508 TFlops after 22 hours; Claude Opus scored 493 TFlops. In another experiment, Step designed training data for preview and increased the accuracy rate from 53.3% to 60% for Qwen3-30b-A3B in AIME24, while consuming fewer token.