AstaBrief 8B
Ai2 open-sourced AstaBrief 8B, a model that turns a research question and retrieved literature excerpts into a cited report, and it runs about 3.5 times faster than Thinking mode.
AstaBrief 8B is built on Qwen3-8B and trained with supervised fine-tuning on 47K examples, then DPO on 6K pairs, all derived from real Asta user queries. It powers Fast mode in Asta, which averages 51.1 seconds per report against 178.5 seconds for the proprietary-model Thinking mode. Weights and training data are released, and an example workflow in the ai2-scholarqa-lib repository lets researchers build reports from their own PDFs.
- Date
- Friday 2 October 2026
- Lab
- Ai2
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Average time per report, Fast mode | 51.1 seconds versus 178.5 seconds for Thinking mode, about 3.5x faster across the full Asta pipeline | company |
| SFT training examples | 47K filtered from about 90K research queries | company |
| DPO pairs | 6K | company |
Search coverage describes an Apache 2.0 license, but the Ai2 page summary did not state the license.
Sources
- allenai.org/blog/astabrief
- hermes-ai.net/news/ai2-s-astabrief-8b-writes-cited-research-reports-3-5x-faster-in-one-pas
This record was partly confirmed: some claims could not be checked on 8 October 2026. How we check