Ai2 releases open AstaBrief model for cited scientific reports

Ai2 announced AstaBrief, an open model for generating cited scientific reports, on October 2, 2026. The release includes the model and its training data, which Ai2 says others can use to examine and build on its approach. AstaBrief 8B takes a research question and retrieved literature excerpts as input, then produces a cited report. Ai2 developed it for research work where outputs need to stay connected to the evidence and researchers need to check the resulting reports.

AstaBrief is offered in Asta’s report-generation feature as Fast mode, alongside a Claude-powered Thinking mode. Across the full Asta pipeline, Fast mode averaged 51.1 seconds per report, compared with 178.5 seconds for Thinking mode, which Ai2 describes as about 3.5 times faster. Ai2 says the comparison reflects proprietary models it tracked, and notes that it has not repeated the full evaluation against current frontier models. The company presents the results as evidence about the training and system choices it tested, rather than a current comparison with those models.

The model starts from Qwen3-8B, with much of the work focused on post-training data, evaluation and the report-generation system around it. Ai2 says development involved tens of thousands of real research queries, citation-focused filtering and preference data. The team chose supervised fine-tuning and direct preference optimization instead of reinforcement-learning-based training, citing a simpler setup that is less costly and easier to manage and debug. Ai2 also redesigned report generation to produce a complete report in one pass from a query and relevant retrieved snippets, rather than creating it section by section.

Ai2 says open weights allow institutions to run AstaBrief on their own infrastructure, which can matter when research questions involve sensitive or unpublished work. The release also includes an example workflow that researchers can adapt to produce reports from their own PDFs. Ai2 says the one-pass approach bypasses snippet summarization and clustering stages used by Thinking mode, and that it found this possible without sacrificing performance. Most of the training and evaluation described in the post was completed in 2025, so the company cautions that its proprietary-model comparisons reflect the frontier at that time.