Ai2 has released AstaBrief 8B, an open-weights language model trained to turn a research question and excerpts from scientific literature into a cited report. It is available in the Asta research platform as a new Fast mode, alongside a Claude-powered Thinking mode, and can also be run on an institution’s own infrastructure.
The model starts from Qwen3-8B and was post-trained with tens of thousands of real research queries, citation-focused filtering and preference data. Ai2 also changed its pipeline to generate a report in one pass instead of composing it section by section. Across the complete Asta workflow, the organization reports an average of 51.1 seconds per Fast-mode report, compared with 178.5 seconds for Thinking mode.
Local deployment matters when a query or document reveals unpublished or sensitive work. Ai2 is releasing model weights, training data and an example workflow for generating reports from private PDFs. It says the model was tuned for relevance, structure and grounding claims in cited evidence.
The comparison has an important limit: most training and evaluation occurred in 2025, and Ai2 has not rerun the full evaluation against current frontier models. The result therefore supports the efficiency of this specialized design, not a broad claim that an 8-billion-parameter model now surpasses the latest general systems.