Recently, Stepfun officially open-sourced a new deep research model called Step-DeepResearch. This model has 32 billion parameters and is dedicated to autonomous information exploration and professional report generation in an open research environment. According to the official introduction, Step-DeepResearch's deep research capabilities are close to top commercial models such as OpenAI's o3-mini and Gemini 2.0 Flash, but its deployment cost is only one-tenth of traditional models, with a single call cost below 0.5 RMB.

image.png

The design concept of Step-DeepResearch is very unique. It breaks down complex research tasks into multiple trainable "atomic capabilities," such as planning, information retrieval, reflection, and cross-validation, thereby achieving closed-loop reflection and dynamic correction. This approach not only enhances the model's adaptability in complex environments but also improves its generalization performance. The training process of the model is carefully designed, from agent mid-training to supervised fine-tuning (SFT) and reinforcement learning (RL), ensuring excellent performance in complex practical applications.

In testing, Step-DeepResearch achieved a high score of 61.4% on the Scale AI Research Rubrics, which is comparable to some larger-scale models such as OpenAI Deep Research and Gemini Deep Research. Additionally, in the expert evaluation of ADR-Bench, Step-DeepResearch's Elo rating was significantly higher than many competitors, demonstrating its strong capabilities in the field of deep research.

To support scientific research workflows, Step-DeepResearch adopts a single-agent architecture based on the ReAct paradigm, featuring a dynamic cycle of reasoning, action, and reflection. Through its internal proprietary toolset, the system can efficiently perform batch web searches, file management, and interactive command execution, providing great convenience for researchers.

github: https://github.com/stepfun-ai/StepDeepResearch

Key points:  

✨ Step-DeepResearch is the latest open-source deep research model from Stepfun, with a parameter scale of 32 billion.  

💡 This model's deep research capabilities are close to top commercial models, but its cost is only one-tenth of traditional models.  

🚀 With a unique training process and dynamic loop architecture, Step-DeepResearch provides efficient support in scientific research.