阶跃星辰发布 Step 5 Preview 开放权重，并公开架构与训练技术博客
Our technical blog, "Inside Step 5 Preview: Architecture and Training for Agentic Work," covers how architecture and training were co-designed for longer, more
阶跃星辰发布 Step 5 Preview 开放权重，BF16 与 FP8 权重已上线 Hugging Face 和 ModelScope，采用 Sparse MoE 架构，总参数 600B、每 token 激活 27B，支持 1M token 上下文窗口和原生文本与视觉输入，面向编码、软件工程、专业知识工作与金融等长程智能体任务。配套技术博客介绍其架构与训练协同设计，包括 92 层窄而深架构用于隐式多跳推理、Sparse GQA 降低百万 token 上下文成本、8-token 块索引将索引与 top-k 开销降至 1/8、关键状态使用 FP32 精度配合细粒度梯度裁剪，以及百万级可验证环境和面向长程任务的 RL 训练。官方还更新了开源智能体框架 Step Code 至 v0.1.4。
推荐理由
AI 评分 82
原文给出 Step 5 Preview 的架构与训练细节，读者可据此判断其面向长程智能体任务的设计取舍。
主题
正文
当前显示原文，中文全文译文尚未就绪。
Our technical blog, "Inside Step 5 Preview: Architecture and Training for Agentic Work," covers how architecture and training were co-designed for longer, more complex agent tasks:
• 92-layer narrow-and-deep architecture for implicit multi-hop reasoning • Sparse GQA to cut million-token context cost, with 8-token block indexing reducing indexing and top-k overhead to 1/8 • FP32 precision on key states plus fine-grained gradient clipping for stable training • Million-scale verifiable environments • RL for long-horizon tasks with context compression built into policy training