AstaBrief 8B: Ai2's Open Model for Cited Science Reports

The Allen Institute for AI (Ai2) released AstaBrief 8B on October 2, 2026, an open-weights language model that turns a research question and a set of retrieved paper excerpts into a structured report with citations. It now powers a new Fast mode inside Ai2's Asta research assistant, and anyone can download the weights to run it on their own hardware.

• AstaBrief 8B is built on Qwen3-8B and released under the Apache 2.0 license, with weights, training data and checkpoints on Hugging Face.
• Ai2 reports an average of 51.1 seconds per report, versus 178.5 seconds for its Claude-powered pipeline, roughly 3.5x faster.
• On the ScholarQA-CS2 test set, the model card lists an average score of 87.0, with 90.5% citation precision and 78.2% citation recall.
• In Asta, 23% of users who tried Fast mode stopped using the slower option entirely, according to Ai2.

What Is AstaBrief 8B?

AstaBrief is a small, specialized model for scientific literature synthesis. Instead of answering from memory, it reads the excerpts a retrieval system hands it and writes a report that cites them. That design keeps answers grounded in real papers, which is exactly what researchers need when they are checking a claim or scoping a new project.

Ai2 trained it on 47,000 supervised fine-tuning examples plus 6,000 DPO preference pairs, drawn from 90,000 filtered real user queries on the Asta platform. The example reports used for training were generated by several frontier models, including Claude 3.5 and 3.7 Sonnet, o3, o4-mini and GPT-4.1. The result is a compact student model that inherits much of their report-writing skill. It follows the same open approach as Ai2's earlier OlmoEarth geospatial platform.

How Good Are AstaBrief's Citations?

The model card's ScholarQA-CS2 numbers show a clear jump over the base model: 87.0 on average, against 77.3 for plain Qwen3-8B. Ai2 also says AstaBrief is competitive with its Claude-powered pipeline and with DR Tulu on answer and citation quality. Those comparisons come from Ai2's own evaluation, so read them as the developer's claims.

Ai2 adds two real-world signals. In a human study, two of three researchers preferred AstaBrief for citation accuracy. In live use, Fast mode earned 84.2% positive feedback, close to the 85.2% for the slower proprietary Thinking mode.

Why Does an Open 8B Research Model Matter?

Speed is the obvious win. A report in under a minute changes how often a scientist will actually ask a question. The bigger story is ownership. An 8 billion parameter open-weight model can run on-premise, which matters for labs handling unpublished results or sensitive data that cannot leave the building. It is also small enough for a well-equipped workstation, much like the Qwen3.8 27B local setup we covered, but at a fraction of the memory.

The weights, data and checkpoints being public also means other teams can study, fine-tune or extend the recipe for their own fields. Follow more open model releases in our AI section.

Sources: Hugging Face Blog: AstaBrief — October 2, 2026; AstaBrief 8B model card — October 2, 2026; Unite.AI — October 2026.