Allen Institute for AI (AI2) has open-sourced AstaBrief, an 8-billion parameter language model purpose-built for generating cited scientific reports from retrieved literature. The release includes model weights, training data, and adaptable workflows, addressing a growing need for open-source alternatives optimized for specialized scientific workflows that demand evidence grounding and citation accuracy. AstaBrief achieves dramatic efficiency gains over AI2's previous Claude-powered system, generating reports in an average 51.1 seconds compared to 178.5 seconds—approximately 3.5 times faster. The team accomplished this performance without sacrificing quality by redesigning the generation pipeline to produce complete reports in a single pass, eliminating expensive intermediate steps like snippet summarization and section-by-section composition. The architectural changes required careful training with real scientific queries, preference data, and citation-focused filtering to maintain grounding in evidence. AI2 opted for a streamlined training approach using supervised fine-tuning and direct preference optimization rather than more complex reinforcement learning methods, prioritizing operational stability and reproducibility. The open-weights model enables institutions to run AstaBrief locally on sensitive or unpublished research data, a critical capability for academic workflows. The initiative advances AI2's broader goal of building open infrastructure for scientific discovery through NSF-funded OMAI, a national program led by AI2 to develop fully open AI systems tailored to research communities.