Open-sourcing AstaBrief, the fast report-generation model in Asta
huggingface ·
Allen Institute for AI has open-sourced AstaBrief, an 8-billion parameter model designed for fast scientific report generation. Built on the Qwen3-8B architecture, the model converts research questions and retrieved literature excerpts into cited reports in a single pass, significantly reducing generation time compared to proprietary alternatives. The training process utilized supervised fine-tuning and direct preference optimization on data derived from real user queries, with rigorous filtering to ensure citation accuracy and evidence grounding. Evaluations indicate that AstaBrief achieves performance competitive with Claude-powered systems while offering nearly four times faster processing speeds in Asta's Fast mode. By releasing both the model weights and training data, the authors aim to enable institutions to run the tool locally for sensitive or unpublished research. This initiative supports broader goals of creating adaptable, open infrastructure for scientific discovery through the NSF OMAI initiative. Early adoption metrics show that users value the speed and quality, with a significant portion continuing to use the Fast mode exclusively.
huggingface