AstaBrief-8B ships open weights, training data and a local example; Ai2 reports improved citation quality on ScholarQA-CS2.
Ai2 opens AstaBrief-8B for fast cited research reports
2026-10-02Verified announcementTRIAL
Ai2 released AstaBrief-8B, an Apache-2.0 model that turns a research question plus retrieved scientific excerpts into a cited report in one pass. It starts from Qwen3-8B, uses supervised fine-tuning and offline DPO, and ships with weights, training datasets and a local PDF example. On the 100-question ScholarQA-CS2 test set, Ai2 reports stronger answer-and-citation scores than its SFT checkpoint and similar aggregate quality to slower multi-step systems.
Availability: Ungated weights, training datasets and local report-generation example
What this means for you
For teams building literature assistants, AstaBrief offers a small, inspectable alternative to multi-step proprietary report pipelines. The reported gains come from Ai2's own retrieval and evaluation setup, so test citation entailment and omitted evidence on your corpus before adoption.
Evidence & related material
Technical details & scores
| Base Model | Qwen3-8B |
|---|---|
| Training | SFT followed by offline DPO; DPO trained on 8 H100 GPUs |
| Data | 39.5K filtered SFT examples and about 6K preference pairs reported by Ai2 |
| Evaluation | ScholarQA-CS2 test average 87.0; citation precision 90.5 and citation recall 78.2, author-reported |
| Availability | Ungated weights, training datasets and local report-generation example |
| License | Apache-2.0 |
Technical impact 3
Engineering value 4
Adoption signal 2
Maturity 3
Verification cost 2
Risk 2