Ai2 has open-sourced AstaBrief, an 8B-parameter model designed to turn a research question and retrieved literature excerpts into a cited scientific report. The model powers the new Fast mode in Asta, the organization's agentic platform for scientific work, alongside a slower Claude-powered Thinking mode. The release includes the model weights, training data, and an example workflow for generating reports from local PDFs.
AstaBrief was trained from Qwen3-8B using supervised fine-tuning and direct preference optimization, a simpler recipe than the reinforcement-learning methods used in some recent open-model work. The authors say they deliberately avoided RL because it can be unstable and expensive, and instead focused on curating high-quality training data from tens of thousands of real research queries, citation-focused filtering, and preference data.
The main efficiency gain comes from generating the full report in one pass, skipping the snippet summarization and clustering stages used by Thinking mode. Across the full Asta pipeline, Fast mode averages 51.1 seconds per report versus 178.5 seconds for Thinking mode—about 3.5× faster—and the authors report nearly an order-of-magnitude reduction compared with the proprietary models they tracked. They also note that open weights are necessary for institutions whose research questions involve sensitive or unpublished work.
One caveat: most training and evaluation was completed in 2025, and the authors have not rerun the full evaluation against today's frontier models. The results, they say, are best read as evidence about the specific training and system design choices tested rather than a current benchmark comparison.