Ohio State's QUEST-35B: The Open-Source Deep Research Agent That Challenges Frontier Systems With Just 32 H100s

A team from Ohio State University's NLP lab, led by Professor Yusu, has released QUEST-35B, an open-source Deep Research agent that demonstrates competitive performance against frontier closed-source systems. The model was trained using approximately 32 NVIDIA H100 GPUs and just 8,000 synthetic samples, marking a significant reduction in the compute and data required for such agents. The team has open-sourced the entire training recipe, code, model weights, and datasets, making it fully reproducible and modifiable by the research community. Benchmark tests show QUEST-35B rivaling systems like those from OpenAI and Google, though remaining gaps include scale of inference compute, long-context retrieval fidelity, and iterative self-correction capabilities found in larger closed models.

This release democratizes access to powerful research agents, enabling smaller teams and institutions to build upon state-of-the-art methods without proprietary infrastructure. The key innovation lies in the efficiency of training—using synthetic data generation to create high-quality research trajectories and leveraging careful instruction fine-tuning. However, the biggest remaining gap between open-source Deep Research agents and frontier closed systems likely lies in the ability to perform multi-step reasoning with very long contexts (100K+ tokens) and integrate real-time web search with citation reliability. Future work may focus on scaling data diversity and inference-time compute budgets to close this gap.