Open-sourcing AstaBrief, the fast report-generation model in Asta

AstaBrief, a report-generation model used within Asta's agentic platform for scientific work, is now being made open-source. The model, called AstaBrief 8B, can generate cited reports from a research question and retrieved literature excerpts. This open model aims to generate high-quality reports faster and at lower costs compared to proprietary models.

AstaBrief uses a simplified training approach, focusing on supervised fine-tuning and direct preference optimization, to reduce the cost and instability often associated with reinforcement-learning-based methods. The training data for AstaBrief was curated from real user queries, ensuring the model learns from actual scientific questions.

The report generation pipeline of AstaBrief generates the entire report in one pass, bypassing the expensive summarization and clustering stages used in other models. This results in a significant speed increase, averaging 51.1 seconds per report compared to 178.5 seconds for a competing model. By making AstaBrief open-source, Asta aims to enable institutions to run the model on their own infrastructure, particularly when dealing with sensitive or unpublished research.

The open weights will also allow other researchers to study, reproduce, and build upon the model's approach. The development of AstaBrief was part of a broader effort by Ai2, a U.S. national initiative, to build open AI infrastructure and models for scientific discovery. The research behind AstaBrief explored how to adapt general-purpose models for scientific work and train new scientific models from scratch, aiming to improve answer quality, relevance, structure, and citation grounding in long-form scientific synthesis.