Ai2 Open-Sources AstaBrief, a Fast 8B Model for Scientific Report Generation
In short: Allen Institute for AI (Ai2) has open-sourced AstaBrief 8B, a model built on Qwen3-8B that turns research questions and retrieved literature excerpts into cited scientific reports. It's now live in Ai2's Asta platform as a 'Fast mode' alongside a Claude-powered 'Thinking mode', generating reports in about 51.1 seconds on average compared to 178.5 seconds for the Claude pipeline — roughly 3.5x faster. Ai2 trained the model using supervised fine-tuning on 47K examples (from 90K filtered real research queries) followed by DPO on about 6K preference pairs, and found that filtering training data for citation density was the single most effective lever for improving grounding quality.
This summary was generated automatically by AI from Hugging Face's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.
What changed?
- 1Open-sourced AstaBrief 8B model weights and training data, built on Qwen3-8B
- 2Generates full cited reports in a single pass instead of section-by-section, cutting generation time
- 3Fast mode averages 51.1s per report vs 178.5s for Claude-powered Thinking mode (~3.5x faster)
- 4Trained via SFT (47K examples) + DPO (~6K preference pairs) rather than RL, for cheaper and more stable training
- 5Evaluated on SQABench-CS2 (200 CS questions) and DeepScholarBench (63 queries); citation-density filtering gave the biggest quality gains
- 6Includes an example local workflow for generating reports from users' own PDFs
| Parameter | Before | Now |
|---|---|---|
| Base model | Not specified | Qwen3-8B |
| Report generation time (avg, full pipeline) | 178.5 seconds (Claude-powered Thinking mode) | 51.1 seconds (AstaBrief Fast mode) |
| Speed vs Thinking mode | 1x (baseline) | ~3.5x faster |
| SFT training examples | Not specified | 47K (from 90K filtered research queries) |
| DPO training examples | Not specified | ~6K |
| Weights/Data availability | Proprietary (Claude 3.5/3.7 Sonnet, o3, o4-mini, GPT-4.1 used for data generation) | Open-sourced model weights and training data |
Why it matters
This shows a trend of smaller open-weight models matching proprietary LLM pipelines on narrow, well-defined tasks (here, cited scientific report writing) at a fraction of the latency and cost, while being self-hostable for sensitive data.
Sources
- Hugging FaceOfficialPrimary sourceOriginal article →„Open-sourcing AstaBrief, the fast report-generation model in Asta“2 Oct 2026, 18:19
- Published by source
- 2 Oct 2026, 18:19
- Found by our system
- 2 Oct 2026, 19:51
- Summary generated
- 2 Oct 2026, 20:07
This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy