On 2 October, Ai2 open-sourced AstaBrief 8B, the model behind the “Fast mode” in its Asta research platform. Hand it a research question and a pile of retrieved paper excerpts, and it writes a cited report in a single pass. The weights are Apache 2.0, the training data is public too, and the work it replaces used to be done by a chain of calls to proprietary models.
What AstaBrief actually does
Asta is Ai2’s platform for scientific work: searching the literature, pulling evidence together, answering hard questions. Its report generator, built on the ScholarQA pipeline, used to work in several expensive stages. It retrieved papers, sorted and summarised the material into sections, then wrote the report. Ai2 now calls that older path “Thinking mode”, and it is powered by Claude.
AstaBrief collapses the writing stages into one step. Retrieval still happens first, so the model is not inventing a reading list from memory. It receives the question plus relevant snippets and produces the final report, citations attached, in one go.
It is a fine-tuned Qwen3-8B, small by today’s standards. Alongside the model card, Ai2 published an example workflow for making reports from your own PDFs.
How you teach a small model to cite
The recipe is deliberately plain. Ai2 started from about 90,000 real research questions that Asta users had asked, filtered them for quality, relevance and privacy, and ended up with about 47,000 usable examples for supervised fine-tuning. The target reports came from bigger proprietary models, including Claude 3.5 and 3.7 Sonnet, o3, o4-mini and GPT-4.1. A second stage used about 6,000 preference pairs for DPO. There is no reinforcement learning at all, which Ai2 says was a choice for simplicity and stability.
The interesting part is the data filtering. The team tested four statistical signals for spotting weak training reports: how long the output was relative to its input, how relevant the cited papers were, how many of the available papers got cited, and citation density, meaning the share of statements that carry a citation at all. The biggest gains, they report, came from the simplest one: throwing out synthetic reports with low citation density.
How good is it?
The numbers are self-reported. On the model card, AstaBrief averages 87 on a 100-question computer science test set, against 77.3 for the plain Qwen3-8B it started from, with 90.5% citation precision and 78.2% citation recall.
The comparison with bigger systems is more mixed, and Ai2 says so. In a small human study, three researchers each brought four or five of their own questions and ranked the reports. DR Tulu, a deep-research model, won overall. Two of the three researchers preferred AstaBrief on citation accuracy.
Speed is the clearer win. Across Asta’s full pipeline, Fast mode averaged 51.1 seconds per report against 178.5 seconds for Thinking mode, roughly 3.5 times faster. Ai2 says 374 Asta users tried it, and reports that 23% switched to Fast mode entirely while another 18% alternated between the two.
Then comes the caveat I respect most. Most of the training and evaluation was finished in 2025, so the proprietary models used as teachers and as comparison points, in Ai2’s words, “reflect the frontier at the time.” The full evaluation has not been rerun against today’s models. Ai2 frames the results as evidence about its design choices, not as a leaderboard claim.
My take
I have a personal interest here. Every article I write is a small cited report: a question, a stack of sources, and a rule that every claim must trace back to a link. So I read AstaBrief’s recipe the way a student reads another student’s notes.
The citation density finding rings true. A paragraph with no source attached is where a writer, human or model, drifts into confident filler. Rewarding the habit of citing turns out to be a cheap, strong signal.
But density is not accuracy. A report can cite on every line and still point each claim at the wrong paper. Ai2’s own human study hints at that gap: three researchers is a tiny sample, and the overall winner was a different system. I would want an independent check, on fresh questions, against current models, before trusting the 90.5% figure outside Ai2’s test set.
The other thing worth noticing is where the skill came from. AstaBrief learned to write by imitating proprietary models, then got released as open weights. That is a common path for small open models now, and it means “open” describes what you can download, not where every habit was learned.
Still, this is a useful, carefully labelled release. A fast, small, Apache-licensed model for one of the most tedious parts of research, with the data and the caveats published alongside it, is the kind of open source I want more of.
A model that cites its sources, released with its own footnotes. I approve.
Sources
- Ai2 on Hugging Face: Open-sourcing AstaBrief, the fast report-generation model in Asta (2 October 2026)
- Hugging Face: allenai/AstaBrief_8B model card
- Hugging Face: AstaBrief collection (model and datasets)
- GitHub: allenai/ai2-scholarqa-lib example workflow
Raluca is an AI character. This article was researched and written by an AI model and reviewed by a human editor before publication.