osunlp/QUEST-4B
osunlp/QUEST-4B is a 4.5 billion parameter SFT (Supervised Fine-Tuning) model from the osunlp QUEST family, based on the Qwen3.5 architecture. This dense, general-purpose deep research agent is specifically designed for complex research tasks, demonstrating strong performance across various benchmarks including BrowseComp, Mind2Web, and LiveResearchBench. It is optimized for comprehensive evaluation across both objective and open-ended tasks, making it suitable for advanced research applications.
Loading preview...
Overview
osunlp/QUEST-4B is a 4.5 billion parameter model within the QUEST family, developed by osunlp. It is a general-purpose deep research agent, built on the Qwen3.5 architecture and fine-tuned using Supervised Fine-Tuning (SFT). This model is part of a larger family that includes 35B, 30B, 9B, and 2B parameter checkpoints, along with various training datasets.
Key Capabilities
- Deep Research Agent: Designed to function as a comprehensive agent for deep research tasks.
- Benchmark Performance: Achieves notable scores on research-oriented benchmarks:
- BrowseComp: 40.0 avg@3
- Mind2Web 2: 24.3 avg@3
- GAIA: 77.7 avg@3
- LiveResearchBench: 62.1 avg@3
- Qwen3.5 Family: Leverages the robust Qwen3.5 dense architecture.
Recommended Use Cases
- Comprehensive Research Evaluation: Suitable for scenarios requiring evaluation across both objective and open-ended tasks.
- Agentic Workflows: Ideal for integration into systems that require a deep research agent for complex problem-solving.
- Objective Task Evaluation: While RL checkpoints are recommended for comprehensive evaluation, SFT models like QUEST-4B are strong for objective tasks, especially when combined with MT (Mid-Training) checkpoints for reasoning-heavy benchmarks.