Allen AI's AstaBrief writes cited research reports from a small open model

The Allen Institute for AI has released AstaBrief, a compact open-weights model that turns a research question plus some literature into a cited report for scientists. It runs as the "Fast mode" inside Asta, next to a heavier "Thinking mode", and the whole point is speed without giving up grounding. AstaBrief produces a report in about 51 seconds, roughly 3.5 times faster than the Claude-powered pipeline it sits beside, which takes around 178 seconds.

What makes the write-up worth reading is how small and reproducible the recipe is. AstaBrief is built on Qwen3-8B and trained with supervised fine-tuning followed by direct preference optimization, using 47,000 SFT examples and about 6,000 DPO pairs drawn from 90,000 real researcher queries. The team found that filtering training data by citation density did more for grounding than any other single change, which is a concrete lesson for anyone building retrieval-heavy assistants. On the SQABench-CS2 and DeepScholarBench evaluations, the 8B model lands competitive with Claude and with DR Tulu on citation precision and answer relevance.

The weights are fully open on Hugging Face, so an institution can run the model on its own hardware instead of sending literature through a proprietary API. Allen AI reports early traction as well: among people who tried it, 29.1% used Fast mode on several days and 23% chose it over Thinking mode every time. You can read the details and the code example in the announcement.

Why it matters

If you work in a lab or hospital that cannot send sensitive literature to an outside API, AstaBrief gives you a self-hostable model that keeps up on citation quality, so grounded literature review no longer forces a privacy tradeoff.