Allen Institute for AI releases open-weight report-generation model AstaBrief 8B

Allen Institute for AI (Ai2) announced the release of AstaBrief 8B, an open-weight model that generates full scientific reports from a user’s research question and retrieved literature excerpts in a single pass. The model is available today within Asta’s "Generate a report" feature as a Fast mode alongside a proprietary Claude-powered Thinking mode, and Ai2 is also open-sourcing the weights and training data.

To train AstaBrief, Ai2 used tens of thousands of real research queries collected from Asta users, citation-focused filtering, and preference data, then applied supervised fine-tuning and direct preference optimization. The team avoided reinforcement-learning-based post-training due to stability and cost concerns, opting for a simpler recipe centered on high-quality training data and a redesigned report-generation pipeline.

AstaBrief writes the full report in one pass rather than section by section, which Ai2 credits for nearly an order-of-magnitude reduction in report generation time. Across the full Asta pipeline, Fast mode averages 51.1 seconds per report compared with 178.5 seconds for Thinking mode, about 3.5× faster.

Ai2 notes that scientific work demands evidence grounding, faithfulness to cited sources, and verifiability—requirements that shaped AstaBrief’s training and evaluation. The institute is exploring broader efforts to adapt open models to scientific workflows through initiatives like NSF OMAI, a U.S. national program led by Ai2 to build open AI infrastructure and models for scientific discovery.

Alongside the model weights, Ai2 released an example workflow that researchers can adapt to generate reports from their own PDFs, providing a starting point for local report generation. The institute emphasizes that open weights enable institutions to run AstaBrief on their own infrastructure, which is necessary when research questions involve sensitive or unpublished work.