Ai2 opens AstaBrief 8B for cited scientific reports
The open-weights model writes reports from research questions and retrieved literature excerpts and now powers a Fast mode in Ai2's Asta service.
An open model for literature synthesis
Ai2 released AstaBrief 8B on October 2 as an open-weights model for generating cited scientific reports. It takes a research question and retrieved literature excerpts, then writes a report in one pass. Ai2 says the model is available in the Fast mode of its Asta research service, alongside an existing Claude-powered Thinking mode. The team also published the model weights and a collection of training data, allowing others to examine or run the report generator on their own infrastructure. Ai2 says the model was adapted from Qwen3-8B using supervised fine-tuning and preference optimization.
Ai2 reports that Fast mode averaged 51.1 seconds per report in its pipeline, compared with 178.5 seconds for Thinking mode. Those are the team’s system measurements, not an independent benchmark of every deployment. Much of its training and evaluation was completed in 2025, and Ai2 says it has not rerun the full comparison against today’s frontier models. The post also notes a limitation beyond citation formatting: a report can cite a relevant study yet overstate what its sample or findings support. AstaBrief’s release makes its approach inspectable, but the published results do not establish that every generated report is accurate.