Ai2 Open-Sources AstaBrief 8B for 3.5 Times Faster Scientific Reports
Ai2 open-sources AstaBrief 8B, a Qwen3-8B report model generating cited scientific syntheses 3.5 times faster than Asta's Thinking mode.
Summary
On October 2, 2026, Ai2 open-sourced AstaBrief 8B, a Qwen3-8B-based model that converts research questions and retrieved excerpts into cited reports. It powers Asta's Fast mode beside Claude-powered Thinking mode, averaging 51.1 seconds per report versus 178.5 seconds, about 3.5 times faster across the pipeline and nearly 10 times faster in generation. Ai2 released its weights, training data and an adaptable PDF workflow through Hugging Face, enabling private deployment behind institutional firewalls for sensitive or unpublished research.
Quality and privacy filters reduced hundreds of thousands of Asta and ScholarQA queries to 90K. Claude 3.5 Sonnet, Claude 3.7 Sonnet, o3, o4-mini and GPT-4.1 produced 47K supervised fine-tuning examples. Separate reports from o3, o4-mini, DeepSeek-V3 and DeepSeek-R1 formed preference pairs; GPT-4.1 and DeepSeek-R1 judges, aligned 95% with humans, unanimously selected 6K direct preference optimization examples. Ai2 chose SFT plus DPO over reinforcement learning and trained one-pass generation that skips summarization, clustering and sectional drafting. Removing low-citation-density reports produced the strongest grounding gains; token ratio, citation relevance, diversity, combined filters and learning-rate sweeps added little.
AstaBrief competed with Claude's pipeline and DR Tulu on the 200-question SQABench-CS2 and 63-query DeepScholarBench. In a 14-question study, three researchers supplied 4 to 5 questions each; DR Tulu won overall, but two preferred AstaBrief for citation accuracy. Most evaluation occurred in 2025 and was not rerun against 2026 frontier models. Among 374 Fast users, 29.1% returned across multiple days and averaged 3.67 report threads; 23% never returned to Thinking, while 18% alternated and used Fast for about 40% of threads. Positive feedback reached 84.2%, versus Thinking's 85.2%. Ai2 plans finer preference learning, RAG with reinforcement learning, multiple turns and tools, more scientific sources, query decomposition, and tests for evidentiary scope, concision and organization.
Positives
- 51.1-second average report generation makes Fast mode about 3.5 times quicker than the 178.5-second Thinking mode pipeline.
- Open weights, training data and a PDF workflow let institutions generate reports locally behind their firewalls.
- 95% human agreement from GPT-4.1 and DeepSeek-R1 judges supported a cleaner 6K-example preference dataset.
- 84.2% positive feedback nearly matched Thinking mode's 85.2% rate among early Asta users.
- Two of three researchers preferred AstaBrief over competing systems on citation accuracy in the 14-question human study.
Risks & concerns
- Most training and evaluation occurred in 2025, with no full comparison against frontier models available in 2026.
- DR Tulu ranked first for overall preference in the three-researcher human evaluation.
- Citation support alone cannot detect whether AstaBrief improperly broadens sample-specific findings or converts descriptions into recommendations.
- 374 Fast mode users and sparse feedback provide limited evidence about broader adoption or long-term research value.
- Initial supervised fine-tuning lagged Claude-powered reports on answer precision and citation quality before additional filtering and DPO.