开源Top2！实测阶跃Step 5 Preview，真有点猛啊…

TL;DR - StepFun’s Step 5 Preview is a 600B-parameter sparse MoE model designed for long-horizon agent tasks while activating only 27B parameters per token. It reportedly ranks second among open models on Artificial Analysis and targets a strong cost-performance tradeoff.

• The model combines a 92-layer “narrow but deep” Transformer architecture with a 1M-token context window to support multi-step reasoning and tool use.
• Sparse MoE, Hybrid Sparse, Sparse GQA, and low-level hardware optimizations aim to reduce the compute and memory costs of long contexts.
• Long-horizon reinforcement learning and context compaction help the model retain task goals across dozens of tool calls and iterative correction loops.
• Hands-on tests produced Blender scenes, 3D and 2D games, and a writing website; outputs were usable after several refinement rounds but still showed occasional layout and detail errors.