Step5preview unveiled as a high-performance, cost-efficient open-source model

Step has officially launched its new flagship model, Step5Preview, designed for real-world Agentic tasks. The model features a sparse MoE architecture with 600B parameters, but only activates 27B at a time. It supports a context window of up to 1 million tokens and processes both text and visual inputs. Step5Preview scored 44 on the Artificial Analysis Intelligence Index, placing it among the top three global open-source models. Its single-task cost is one-eighth that of Claude Opus5, offering stronger capabilities at a lower price. The model excels in high-difficulty benchmarks, ranking second only to GPT-6Astra or Claude Opus5. It also performed well in programming tasks, demonstrating strong success rates and cross-scenario stability. In a notable demonstration, Step5Preview transformed an ESP32-S3 board into a Vibe Coding keyboard with Bluetooth and voice input capabilities. Long-term tasks further showcased its strength, with Step5Preview achieving 508 TFLOPS in GPU kernel optimization, surpassing Claude Opus5. The model’s weights will be fully released on October 15th.