Hugging Face AI Updates: October 3, 2026

1. Ai2 Open-Sources AstaBrief, an 8B Model for Cited Scientific Reports

Hugging Face. Ai2 published AstaBrief on the Hugging Face blog, an 8-billion-parameter open-weights model built on Qwen3-8B that generates cited scientific reports from a research question and retrieved literature excerpts. The model was trained with supervised fine-tuning followed by direct preference optimization rather than reinforcement learning, using 47,000 SFT examples and about 6,000 DPO preference pairs derived from 90,000 filtered real user queries. In the full Asta pipeline, AstaBrief’s Fast mode averages 51.1 seconds per report versus 178.5 seconds for Thinking mode, about 3.5x faster, and Ai2 reports competitive results on SQABench-CS2, DeepScholarBench, and pairwise comparisons against Claude-powered systems. Weights are available at allenai/AstaBrief_8B alongside the training datasets, and the model now powers report generation in Asta. Source