inclusionAI/DR-Venus-4B-SFT
DR-Venus-4B-SFT is a 4 billion parameter deep research agent developed by inclusionAI, fine-tuned from Qwen/Qwen3-4B-Thinking-2507. It specializes in establishing stable long-horizon agentic behavior, including reasoning, tool use, evidence collection, and final answer synthesis. This model is optimized for deep research agents with long-horizon tool use and open-domain information seeking using search and visit tools. It serves as a strong baseline and initialization checkpoint for more advanced agentic models, outperforming other small agents on deep research benchmarks.
Loading preview...
DR-Venus-4B-SFT: A Deep Research Agent
DR-Venus-4B-SFT is a 4 billion parameter deep research agent developed by inclusionAI, built upon the Qwen/Qwen3-4B-Thinking-2507 base model. It is the supervised initialization checkpoint for the DR-Venus project, designed to enable stable, long-horizon agentic behavior.
Key Capabilities
- Long-Horizon Agentic Behavior: Excels in complex tasks requiring multi-step reasoning, tool use, and evidence synthesis.
- Tool-Augmented Research: Specifically trained for open-domain information seeking using
searchandvisittools. - Open-Data Training: Fine-tuned on cleaned open-data agent trajectories from REDSearcher, ensuring transparency and reproducibility.
- Strong Baseline Performance: Achieves competitive results on deep research benchmarks like BrowseComp, GAIA (Text-Only), xBench-DS, and DeepSearchQA, often outperforming other agents under 9B parameters.
Intended Use Cases
- Deep Research Agents: Ideal for applications requiring extensive, tool-driven research.
- Information Seeking: Suited for tasks involving web search and data retrieval.
- Agentic Checkpoint Initialization: Serves as a robust starting point for further reinforcement learning (RL) in agent development.
- DR-Venus Inference Pipeline: Designed for deployment within the official DR-Venus inference pipeline for optimal performance and tool protocol adherence.