Ai2 releases AstaBrief 8B for fast, cited scientific report generation

Ai2 has released AstaBrief, an 8B open-weight model designed for a narrow but demanding task: turning a research question plus retrieved scientific excerpts into a cited report. The project is notable because it treats source grounding as part of the model’s job rather than an optional layer added after generation. Ai2 says AstaBrief is now used in Asta’s Fast mode and is being released with the training data so that researchers can inspect, reproduce and adapt the approach.

The model starts from Qwen3-8B and was post-trained with supervised fine-tuning and direct preference optimization. Ai2 describes a training pipeline built from real research queries, citation-focused filtering and preference pairs. After filtering, the team reports roughly 47,000 SFT examples and about 6,000 DPO pairs. The important methodological point is that Ai2 did not present one judge model or one synthetic generator as ground truth; it used multiple model sources and agreement filters to reduce noise.

For open-weight users, the practical value is local control over a report-generation component that would otherwise depend on a proprietary API. Ai2 reports that Fast mode averaged 51.1 seconds per report across its pipeline, compared with 178.5 seconds for its Claude-powered Thinking mode. The source also explicitly warns that much of the evaluation was completed earlier and should not be read as a current frontier ranking. For Open Model Weights, that distinction matters: the release is interesting because the weights and training artifacts are available, while benchmark and speed claims remain attributed to the publisher until independently reproduced.