Ai2 Open-Sources AstaBrief 8B for Faster Cited Scientific Reports

AstaBrief was built from Qwen3-8B, with most effort going into post-training data, evaluation and the surrounding report-generation scaffolding, the blog post says. Ai2 said it wanted to test whether a small, open model trained specifically for scientific report generation could match the report quality of proprietary models while reducing generation time and serving costs.

Developing the model required tens of thousands of real research queries, citation-focused filtering, preference data and a redesigned pipeline that writes the full report in one pass rather than section by section. The blog post says this produced nearly an order-of-magnitude reduction in report generation time compared with the proprietary models Ai2 tracked. Across the full Asta pipeline, Fast mode averaged 51.1 seconds per report, compared with 178.5 seconds for Thinking mode, about 3.5 times faster.

Ai2 says open weights will let institutions run AstaBrief on their own infrastructure, which it describes as necessary when research questions reveal sensitive or unpublished work. Alongside the weights, Ai2 is releasing an example workflow that researchers can adapt to create reports from their own PDFs, providing a starting point for local report generation.

For training, Ai2 considered reinforcement-learning methods, including recent work such as DR Tulu, which has shown that RL can improve long-form report generation for open-weights models, especially when judge models are involved in the training loop. The team ultimately focused on a simpler recipe built around supervised fine-tuning and direct preference optimization, citing RL's instability and expense and a desire for a setup that is cheaper, easier to debug and easier to iterate on. That made training-data quality especially important: rather than relying on a more complex optimization method to compensate for noisy examples, Ai2 spent much of the project generating, selecting and filtering examples that demonstrated the report-writing behavior it wanted.

Ai2 says the goal was an open-weights model with the qualities that matter most for long-form scientific synthesis: answer quality, relevance, structure and citation grounding. The blog post says most of the training and evaluation described was completed in 2025, so the proprietary models used to generate training data and as comparison points reflect the frontier at the time. Ai2 has not rerun the full evaluation against today's frontier models and says the results are best read as evidence about the particular training and system design choices it tested.

The work is part of Ai2's broader exploration of adapting general-purpose models for science and training new scientific models from scratch. Through NSF OMAI, a U.S. national initiative led by Ai2 to build fully open AI infrastructure and models for scientific discovery, Ai2 researchers are working directly with scientific communities to understand what they need from future open models and where today's general-purpose models fall short, including how needs differ across scientific fields and workflows.

Editor's Summary

Ai2 has open-sourced AstaBrief 8B, a Qwen3-8B-based model that generates cited scientific reports in Asta's Fast mode, and released its training data and an example local workflow. The blog post reports faster generation than the proprietary models it tracked, while noting that most training and evaluation was completed in 2025 and has not been rerun against current frontier models.