Ai2 releases AstaBrief 8B for faster cited research reports
Ai2 opens its scientific report model and training data, with faster generation, older evaluation baselines and different licenses for weights and data.
Ai2 has released AstaBrief 8B, a downloadable model that turns a research question and supplied scientific excerpts into a report with citations. The October 2 announcement also presents it as Fast mode in Asta, alongside the existing Thinking mode powered by Claude.
Ai2 reports an average of 51.1 seconds per report across the full Fast mode pipeline, compared with 178.5 seconds for Thinking mode. That is about 3.5 times faster in its measurements. Most development and evaluation happened in 2025, and Ai2 has not repeated the full comparison against current frontier models.
A faster route through the evidence
AstaBrief generates the report in one pass from retrieved excerpts. It skips the intermediate summarization and grouping stages used by Thinking mode. The speed result therefore reflects changes to the model and the surrounding workflow. Ai2 also provides an example for building reports from local PDFs.
For a research team, the practical attraction is a shorter wait before reviewing a first draft. The useful test is whether that draft reduces the work required to trace claims back to evidence. Fast prose alone does not settle that question.
What the released model actually measures
The model card identifies Qwen3 8B as the base and reports gains after supervised training and direct preference optimization. On the 100 question ScholarQA CS2 test split, its average across four measures rises from 77.3 for the base model to 87.0 for AstaBrief.
The improvement is uneven. Citation precision rises from 76.2 to 90.5 and citation recall from 64.6 to 78.2. Answer precision falls from 90.6 to 89.0. Those measures address different parts of a report, so the average should not be read as a single probability that an answer is correct.
The comparison also shows a tradeoff across benchmarks. AstaBrief scores 53.50 on DeepScholarBench, below Asta ScholarQA at 60.25 and DR Tulu 8B at 56.26. The card recommends retaining its supplied prompt format because other formats can produce inconsistent results.
Training data puts citations under pressure
The supervised training dataset documents a pool of 90,000 filtered research queries from users who opted into data sharing. The queries came from OpenScholar and Asta ScholarQA through June 2025. Filtering removed bot and test traffic, very short prompts, non English queries, requests outside science and personal medical information.
Ai2 initially retained 47,000 generated report examples. It then removed reports where fewer than a quarter of statements had citations, leaving 39,500 examples. That distinction matters when comparing the announcement with the released dataset. The smaller number reflects an additional citation density filter.
The separate preference dataset contains about 6,600 examples. Each pairs two reports, with preferences retained when GPT 4.1 and DeepSeek R1 agreed on the winner. This makes the training judgments inspectable, while leaving them dependent on the models used to create and evaluate the reports.
Open weights and training data have different terms
The model weights carry an Apache 2.0 license. The card describes research and educational use under Ai2 responsible use guidance.
Both released training datasets carry Creative Commons Attribution NonCommercial 4.0 terms. Their documentation also says synthetic outputs remain subject to the respective providers’ terms. Access to the weights and permission to reuse the training data are separate questions for organizations considering commercial development.
A useful starting point for careful research
Ai2 cautions that citation grounding does not fully measure whether a report preserves the scope and strength of a study. Its next evaluations will need to examine that distinction more closely.
The release gives institutions a concrete model and training recipe to inspect. A sensible evaluation would compare reports on the team’s own literature, check whether citations support each claim and include the time spent correcting errors.
For another approach to making scientific AI work inspectable, read ByteForward’s coverage of the BootLoops research toolkit.
Illustration is original AI generated artwork created for ByteForward. It represents linked research evidence and does not depict the Asta interface.
