Step 5 preview model launches with 1 million token context window

On October 15, 2026, Jieyu released the Step 5 Preview model weights, including BF16 and FP8 weights, technical blogs, and open-source development tools. The Step 5 Preview is designed for long-range agent tasks, utilizing a sparse MoE architecture with a total parameter count of 600B and an activation parameter of 27B per token. It supports a 1 million token context window and natively handles text and visual inputs, excelling in AI programming, software engineering, professional work, and finance. Recently, it ranked second globally in OpenRouter weekly and daily usage metrics. The model weights are now available on Hugging Face and ModelScope, with support for various chip platforms and inference frameworks. Developers can deploy compatible OpenAI API services using vLLM or SGLang, or perform offline inference with Transformers, integrating the model into their data, tools, and systems. The technical blog highlights architectural and training optimizations for longer, more complex agent tasks, including Sparse GQA, stable training, a million-level verifiable environment, and reinforcement learning for long-range tasks.