AstaBrief 8B: Ai2 open-weights cited scientific reports for Asta Fast mode

Ai2’s Hugging Face blog on October 2, 2026 open-sources AstaBrief 8B—a model that turns a research question plus retrieved literature excerpts into a cited scientific report (blog, Kyle Wiggers / Ai2Comms). Weights and training data are released; the model card states Apache-2.0.

What it does

AstaBrief is built for cited scientific report generation, not general chat. In Asta, it powers Generate a report → Fast mode today, alongside a Claude-powered Thinking mode. Ai2 says Fast mode writes the full report in one pass given the query and retrieved snippets, rather than Thinking’s section-by-section path with heavier snippet summarization and clustering.

Training recipe

Ai2 started from Qwen3-8B, then post-trained with supervised fine-tuning (SFT) and direct preference optimization (DPO). Reinforcement learning was considered and not used for this release. The card notes the DPO checkpoint builds on an SFT sibling and preference pairs over report alternatives from multi-model synthetic targets.

Speed and evals (Ai2-attributed)

Ai2 reports full Asta pipeline averages of about 51.1 seconds per report in Fast mode versus about 178.5 seconds in Thinking mode (~3.5×)—vendor pipeline times, not an independent newsroom bench. The blog’s own caveat: most training and evaluation finished in 2025, proprietary baselines reflect that era, and Ai2 has not rerun the full eval against today’s frontier models. Read the tables as approach and system-design validation, not a current frontier ranking. Early product usage notes (374 Asta users who tried Fast, with retention and feedback figures in the blog) stay Ai2-attributed early signals.

Who this is not

This is a specialized Asta report model—not Olmo-core 3 training-stack news, not a general chat-model GA, and not a substitute claim that Fast mode replaces Thinking for every research workflow.

Who should care

Teams who want open weights for one-pass cited scientific reports—or who already use Asta Generate-a-report—should start at the announcement and allenai/AstaBrief_8B. Treat latency and 2025-era eval numbers as Ai2’s stated evidence about the design they tested, not as today’s leaderboard.