source: Hugging Face Blog: Open-sourcing AstaBrief, the fast report-generation model in Asta
level: technical
ai2 open-sourced astabrief 8b, a model that turns a research question and retrieved literature excerpts into a cited report. it is available in asta's generate a report feature as fast mode, alongside claude-powered thinking mode. the model was trained on tens of thousands of real research queries, with citation-focused filtering and preference data. it writes the full report in one pass rather than section by section, reducing generation time from 178.5 seconds to 51.1 seconds on average.
the training data came from 90k filtered research queries, yielding 47k supervised fine-tuning examples and 6k direct preference optimization pairs. evaluation on sqabench-cs2, a 200-question computer science benchmark, showed astabrief matched the claude-powered pipeline on rubric score and citation precision. however, the evaluation was completed in 2025, so the proprietary models used for comparison are not current frontier models. the authors note that citation support alone does not guarantee scientific faithfulness, as models can overstate claims.
astabrief is based on qwen3-8b and uses supervised fine-tuning plus direct preference optimization, avoiding more expensive reinforcement learning. the open weights let institutions run the model locally, which matters when research questions involve sensitive or unpublished work. ai2 also released an example workflow for generating reports from local pdfs. this work is part of nsf omai, a national initiative to build open ai infrastructure for scientific discovery.
why it matters: open-weights report generation models let researchers run fast, cited literature synthesis on their own infrastructure, protecting sensitive data and reducing costs.
source: Hugging Face Blog: Open-sourcing AstaBrief, the fast report-generation model in Asta