A small open model writes a sourced literature report
Ai2’s AstaBrief 8B takes a research question and retrieved literature excerpts, then produces a structured report with citations. The team says it redesigned its Asta pipeline so the model writes the full report in one pass rather than first summarizing and clustering snippets. The model weights and training data are being released.
The result is meant for early synthesis and iteration: a researcher can get a report quickly, inspect its sources and then refine the question. Open weights also make it possible for institutions to run the model on their own infrastructure when research questions involve sensitive or unpublished work.
The speed number is about the full workflow
Ai2 reports that Fast mode averaged 51.1 seconds per report, compared with 178.5 seconds for its Claude-powered Thinking mode, or about 3.5 times faster. That measures the full Asta pipeline, not just raw token generation.
The comparison has a date limitation: Ai2 says most of the evaluation and proprietary model comparisons were done in 2025 and were not rerun against the frontier models available today. The result shows what the team achieved in that setup, not a current ranking of scientific models.
What the release gives builders
The write-up describes the data and post-training recipe and links to an example PDF-based workflow. Ai2 says its training used filtered real research queries and citation-focused examples. Builders can inspect the released weights and data rather than treating the report generator as a closed service.
For scientific use, the central question remains whether each claim is supported by the cited paper and whether the report leaves out conflicting evidence. Fast, open report generation is useful when it shortens literature triage; it does not remove the need to read and verify the underlying sources.
Source published October 2026. Coverage is based on the maker’s announcement and demonstration.