Open-sourcing AstaBrief, the fast report-generation model in Asta

Ai2 has open-sourced AstaBrief 8B, the report-generation model behind the Fast mode in its Asta scientific research platform. It takes a research question plus retrieved literature excerpts as input and writes a full cited report in a single pass — no section-by-section generation. Weights and the training data are both being released.

It's built on Qwen3-8B, with most of the engineering effort going into post-training data, evaluation, and the report-generation scaffolding rather than the base model.

Performance and setup

• Across the full Asta pipeline, Fast mode averages 51.1 seconds per report, compared with 178.5 seconds for the Claude-powered Thinking mode — about 3.5× faster, described as nearly an order-of-magnitude reduction in generation time vs the proprietary models they tracked.
• AstaBrief is live today in Asta's Generate a report feature as Fast mode, running alongside Thinking mode.
• Ai2 also ships an example workflow that researchers can adapt to generate reports from their own PDFs, aimed at local report generation.
• Open weights let institutions run it on their own infrastructure, which matters when research questions involve sensitive or unpublished work.

Training recipe: SFT + DPO, not RL

Ai2 explicitly chose a simpler recipe: supervised fine-tuning plus direct preference optimization. They point to their earlier DR Tulu work showing RL can improve long-form report generation for open-weights models, but say RL-based training is unstable and expensive. They wanted a setup that's cheaper to run, easier to debug, and easier to iterate on — which pushed the emphasis onto training data quality.

Building AstaBrief required tens of thousands of real research queries, citation-focused filtering, preference data, and a redesigned pipeline that writes the report in one pass. The model's stated goals are answer quality, relevance, structure, and citation grounding — keeping answers tied to what the evidence actually supports instead of silently broadening a study's conclusions.

Caveats

Most training and evaluation was completed in 2025, so the proprietary models used to generate training data and as comparison points reflect the frontier at that time. Ai2 has not rerun the full evaluation against current frontier models, and frames the results as evidence about the specific training and system design choices they tested. The work also ties into NSF OMAI, the U.S. initiative led by Ai2 to build open AI infrastructure and models for scientific discovery.

If you run local models and want cited long-form synthesis over your own PDFs, the released example workflow is the place to start.