QUEST: Training Frontier Deep Research Agents with Fully Synthetic Tasks

Researchers introduced QUEST, an open family of deep research agents ranging from 2B to 35B parameters.

The models were trained with fully synthetic tasks, supervised fine-tuning, and reinforcement learning, plus built-in context management.

On eight deep research benchmarks, QUEST performs strongly and ranks among the best open-weight agents.

The result suggests open models can approach or surpass proprietary systems on complex research, citation grounding, and report synthesis.