Ai2 has released AstaBrief 8B, a model that turns a research question and retrieved scientific literature excerpts into a report with citations. It already powers Fast mode in Asta’s Generate a report tool, alongside Claude-powered Thinking mode. The release includes model weights and training data, so researchers can run the report writer on their own infrastructure.
Across the full Asta workflow, Ai2 reports an average of 51.1 seconds per report in Fast mode, compared with 178.5 seconds in Thinking mode. That is about 3.5 times faster. Most of the training and evaluation took place in 2025, and Ai2 says it has not rerun the full evaluation against current frontier models.
What researchers can download
AstaBrief starts from Qwen3-8B. Ai2 trained it through supervised fine-tuning, which teaches a model from example reports, followed by direct preference optimization, which teaches it to favor one report over another.
The model card lists an Apache 2.0 license, an intermediate fine-tuned checkpoint and the preference dataset. It also documents training settings and inference code. Ai2’s release collection gathers the datasets and checkpoints from the training process.
Ai2 also links an example workflow for generating reports from researchers’ own PDFs. The downloadable model takes a question and literature excerpts as input. A local deployment still needs a way to supply those excerpts. The model card recommends the prompt format used during training and warns that a different format may produce inconsistent results.
Why Fast mode takes less time
Thinking mode summarizes and groups retrieved excerpts before writing the report section by section. AstaBrief generates the full report in one pass. Its speed result therefore reflects both a smaller model and a simpler report-generation workflow.
Ai2 built the training material from a filtered pool of 90,000 research queries. Reports generated by proprietary models produced 47,000 usable supervised training examples. A separate set of about 6,000 report pairs supplied preference training, with GPT-4.1 and DeepSeek-R1 judging the pairs. Ai2 kept pairs where both judges agreed.
Filtering out training reports with too few cited statements produced the strongest gains among the data filters Ai2 tested. More aggressive filtering and combinations of filters did not add meaningful gains.
What the evaluations establish
Ai2 evaluated coverage, relevance and citation support on computer science research questions, with a separate benchmark for long-form research synthesis. Its model card reports citation precision of 90.5 for AstaBrief, compared with 76.2 for base Qwen3-8B. Citation recall rose from 64.6 to 78.2, while answer precision fell from 90.6 to 89.0 on the same test set.
The human comparison was small. Three researchers contributed 14 questions in total. DR Tulu won on overall preference, while two researchers preferred AstaBrief on citation accuracy. Ai2 also acknowledges that citation support alone can miss a report that broadens a study’s conclusion beyond the population or conditions it examined.
Sources
- Ai2: Open-sourcing AstaBrief, the fast report-generation model in Asta. Source of the workflow timings, training account and lead artwork, published by Ai2.
- Ai2: AstaBrief_8B model card
- Ai2: AstaBrief release collection
