Summary

Step-DeepResearch is an end-to-end deep research agent model designed for autonomous information exploration and professional report generation in open-ended research scenarios. It targets researchers and professionals seeking cost-effective, autonomous solutions for complex information gathering and analysis. The system offers deep research capabilities, closed-loop reflection, and dynamic correction within a single inference pass, significantly reducing operational costs.

How It Works

The agent employs a single-agent architecture based on the ReAct paradigm, orchestrating reasoning, action, and reflection. It decomposes complex research tasks into trainable atomic capabilities—planning, information seeking, reflection, cross-validation, and professional report generation. This approach ensures closed-loop reflection and dynamic correction within a single inference pass. A progressive training pipeline, spanning Agentic Mid-Training, Supervised Fine-Tuning (SFT), and Reinforcement Learning (RL), shifts the training objective from simple token prediction to deciding the next atomic action, thereby enhancing adaptive capabilities and generalization performance. It integrates a proprietary toolset with local implementations, including batch_web_surfer for batch web searches, file operations, todo task management, and shell for interactive command execution, supporting a complete research workflow.

Quick Start & Requirements

• Primary Install: Backend: uv sync or pip install -e .; Frontend: npm install or yarn install.
• Prerequisites: Python >= 3.10, Node.js >= 18, npm/yarn. Requires StepFun API keys for model and search services.
• Setup: Configure environment variables (MODEL_PROVIDER, MODEL_BASE, STEP_MODEL_API_KEY, STEP_SEARCH_API_BASE, STEP_SEARCH_API_KEY).
• Links: Technical report available. Demo UI and offline runner instructions provided. Beta API access via a form.

Highlighted Details

• With only 32B parameters, Step-DeepResearch achieves 61.4% on Scale AI Research Rubrics, matching the performance of OpenAI Deep Research and Gemini Deep Research.
• In expert human evaluations on the ADR-Bench, its Elo score significantly outperforms larger models including DeepSeek-v3.2 and GLM-4.6, and rivals top-tier closed-source models.
• Offers superior cost-effectiveness, providing expert-level research capabilities with extremely low deployment and inference costs, positioning it at the high-efficiency frontier.
• Atomic Capability Integration ensures closed-loop reflection and dynamic correction within a single inference pass by decomposing tasks into trainable atomic capabilities.

Maintenance & Community

Users can join a group chat for updates on beta API application status and project developments. Specific details on contributors, sponsorships, or a public roadmap are not provided in the README.

Licensing & Compatibility

The code is licensed under the Apache 2.0 License, which generally permits commercial use and integration into closed-source projects.

Limitations & Caveats

The system relies on proprietary StepFun API keys for its core functionalities. Beta API access is mentioned, suggesting potential ongoing development or limited availability. Specific limitations or known issues are not detailed in the provided README.