Ai2 has released AstaBrief 8B, an open-source model that produces scientifically sourced summaries 3.5 times faster

Trained on 90,000 real queries from researchers, Ai2’s small model generates a cited report in a single pass and powers the new ‘Fast’ mode on the Asta platform.

The Allen Institute for AI (Ai2) has released the weights and training data of AstaBrief 8B, a model responsible for transforming a research question and excerpts from scientific articles into a structured, citation-based report. It now equips the reporting function of its Asta platform, in "Fast" mode, alongside a "Thinking" mode that relies on Claude.

A deliberately simple recipe

Built from Qwen3-8B, the model was refined by supervised learning and then by direct preference optimization (DPO), without reinforcement learning, which was considered more unstable and costly. The effort focused on the data: 90,000 real queries from researchers, filtered for quality and confidentiality, yielded 47,000 usable examples, supplemented by pairs of reports separated by two judge models.

The main gain is speed: AstaBrief writes the full report in one piece instead of building it section by section. Over the entire pipeline, a report takes an average of 51.1 seconds compared to 178.5 in Thinking mode. Ai2 specifies that most of the evaluations date from 2025 and have not been redone in the face of current cutting-edge models. The open weights, on the other hand, allow institutions to run it on their own servers, an asset for sensitive or unpublished work.