inclusionAI/DR-Venus-4B-RL
DR-Venus-4B-RL by inclusionAI is a 4 billion parameter reinforcement-learned deep research agent, built on Qwen3-4B-Thinking-2507. It is specifically designed for long-horizon web research, featuring explicit tool use with 'search' and 'visit' functions, evidence collection, and answer generation. This model excels at improving execution reliability in multi-step retrieval and browsing tasks, trained with IGPO-style information gain rewards and format-aware turn-level supervision.
Loading preview...
DR-Venus-4B-RL: A Deep Research Agent
DR-Venus-4B-RL is a 4 billion parameter model developed by inclusionAI, specifically engineered for long-horizon web research and evidence-grounded question answering. It is a reinforcement-learned checkpoint, building upon the inclusionAI/DR-Venus-4B-SFT model, and uses Qwen/Qwen3-4B-Thinking-2507 as its base.
Key Capabilities & Training
This model's core strength lies in its agentic capabilities, trained with an advanced IGPO-style agentic RL algorithm. It leverages information gain rewards and format-aware turn-level supervision to enhance execution reliability over long tool-use trajectories. The training incorporates a maximum rollout horizon of 200 interaction steps and supports a maximum context length of 256K, enabling extensive multi-turn interactions with search and visit tools.
Performance Highlights
DR-Venus-4B-RL demonstrates significant improvements over its SFT counterpart and other small models (under 9B parameters) on various deep research benchmarks. It shows gains in:
- BrowseComp: +2.3
- BrowseComp-ZH: +2.0
- xBench-DS-2505: +5.7
- xBench-DS-2510: +5.4
- DeepSearchQA: +1.9
These improvements are attributed to better formatting accuracy, more reliable tool use, and enhanced long-horizon execution stability.
Intended Use Cases
- Long-horizon deep research with tool-augmented reasoning.
- Improving execution reliability in complex, multi-step tasks.
- Evidence-grounded answering using
searchandvisittools. - Deployment within the official DR-Venus inference pipeline.
It is not primarily optimized for plain chat or generic short-context instruction following without tools.