Ai2 released AstaBrief 8B on October 2, with Apache 2.0 weights and training data. The release links code for adapting reports to your own PDFs. It starts from Qwen3-8B and writes a cited scientific report from a question plus retrieved excerpts. Ai2 reports 51.1 seconds for its full Fast pipeline versus 178.5 seconds for Thinking mode, about 3.5 times faster. Much of the evaluation was completed in 2025; Ai2 explicitly has not rerun the full comparison against today’s frontier.
The model card’s citation precision is 90.5% and citation recall 78.2% on its computer-science test set. Those are different properties: citations can support the statements beside them while other statements remain uncited. Neither statistic guarantees that the report preserves a paper’s population, time frame or strength of conclusion. The linked code takes already retrieved references and metadata, then generates and parses a report. Its parent project defaults to Semantic Scholar’s public API. An institution seeking a private workflow would need to configure local retrieval and model serving; local weights alone do not establish confidentiality.
Release and limitations · Model, prompt and evaluation table · Report-generation code.