osunlp/QUEST-9B
QUEST-9B is a 9 billion parameter SFT (Supervised Fine-Tuned) model from the osunlp team, built on the Qwen3.5 family architecture. It functions as a general-purpose deep research agent, specifically optimized for complex objective tasks and excelling in benchmarks like GAIA and LiveResearchBench. This model is designed for applications requiring advanced reasoning and information retrieval capabilities over open-ended tasks.
Loading preview...
QUEST-9B: A Deep Research Agent
QUEST-9B is a 9 billion parameter model developed by osunlp, part of the Qwen3.5 family, and is specifically supervised fine-tuned (SFT) to act as a general-purpose deep research agent. It is designed to handle complex objective tasks, demonstrating strong performance across various benchmarks.
Key Capabilities & Performance
This model shows notable performance in benchmarks related to research and browsing tasks:
- BrowseComp: Achieves an average score of 45.4.
- Mind2Web 2: Scores 24.4 on average.
- GAIA: Demonstrates strong performance with an average score of 78.6.
- LiveResearchBench: Achieves an average score of 63.5.
These results indicate its proficiency in tasks requiring information gathering, web interaction, and complex problem-solving.
Model Family and Selection
QUEST-9B is part of a larger family of QUEST models, including 35B, 30B, 9B, 4B, and 2B parameter variants. For objective tasks that are reasoning-heavy, the MT+SFT checkpoints within the QUEST family are recommended. For a more comprehensive evaluation across both objective and open-ended tasks, the RL checkpoints are suggested. QUEST-9B is released under the Apache License 2.0.