The Allen Institute for AI has released AstaBrief 8B, the open-weights model behind the fast report-generation mode in its Asta research assistant. It is built on Qwen3-8B, and the release includes training data as well as weights: roughly 47,000 supervised fine-tuning examples drawn from filtered real user queries, and about 6,000 preference pairs, with citation-focused filtering applied. Example code for generating reports from local PDFs is published alongside it.
Evaluation is on SQABench-CS2, a set of 200 computer science research questions scored for rubric quality, answer precision, and citation precision and recall, with secondary results on the 63-query DeepScholarBench, pairwise judgements and a small human study.
The caveats are the reason this release is worth noting at all, because Ai2 states them itself. Most of the training and evaluation was completed in 2025. The full evaluation has not been rerun against today’s frontier models, and the institute says the results are best read as evidence about particular training and system design choices rather than as a comparison with the current state of the art. The human study involved three researchers and fourteen questions.
The usage figures are small and reported as such: of 374 Asta users, 29.1% have used the fast mode on two or more days.
Source: Allen Institute for AI, Open-sourcing AstaBrief, the fast report-generation model in Asta.