Ai2 has released AstaBrief 8B, an open-weight model that turns a research question and retrieved literature excerpts into a cited report. In Ai2’s Asta platform, it powers a new Fast mode that averaged 51.1 seconds per report across the full pipeline, compared with 178.5 seconds for the Claude-powered Thinking mode. That is about 3.5 times faster in Ai2’s own test.
The speedup comes partly from changing the workflow, not just shrinking the model. AstaBrief starts from Qwen3-8B and generates a full report in one pass from retrieved passages. Ai2 says this bypasses the snippet-summarization and clustering stages used by Thinking mode. The model does not search the literature on its own: its input includes the question and material retrieved upstream.
Ai2 trained AstaBrief first on example reports, then on pairs of reports ranked by how well they followed the evidence; the team calls the second step direct preference optimization. From a privacy- and quality-filtered pool of 90,000 research queries, it produced 47,000 training reports and about 6,000 preference examples. The release includes the model weights and preference data, giving other teams material to inspect and adapt rather than access only to a hosted feature.
Ai2’s main test, SQABench-CS2, contains 200 user-written computer-science research questions. It tracks report coverage, relevance, citation precision and citation recall: whether each citation supports the claim beside it, and whether the report’s claims are backed by citations. Ai2 says citation scores alone cannot show whether a report preserves the scope and limits of the original studies, a separate problem the team identifies for future evaluation.
Ai2 says most of the training and evaluation happened in 2025, and it has not rerun the full comparison against today’s leading models. The reported speed is also a Fast-versus-Thinking comparison within Ai2’s own service, not an independent benchmark against other report systems. For teams considering local research assistants, the release offers a concrete starting point for testing evidence-linked reports on their own infrastructure. The next question is whether the latency and citation results hold across other fields, retrieval systems and current model baselines.
Sources
- Ai2: Open-sourcing AstaBrief, the fast report-generation model in Asta (Oct. 2, 2026)
- AstaBrief DPO preference dataset
- Image: “The life of Research paper,” by Gn.zhigerbayeva, licensed under CC BY-SA 4.0.
