Ai2 releases AstaBrief, an open model that writes cited science reports

The AI research institute Ai2 (Allen Institute for AI) released AstaBrief 8B on October 2, 2026. The 8-billion-parameter model takes a research question and excerpts retrieved from scientific papers and writes a report with citations. It is available under the Apache 2.0 license, and its weights can be downloaded, so you can run AstaBrief locally on your own machine.

AstaBrief and Asta

AstaBrief is part of Asta, Ai2's agentic platform for scientific work. According to Ai2's post on Hugging Face, the model powers Asta's new Fast mode, which can be tried at asta.allen.ai alongside the existing Thinking mode. Ai2 previously produced these reports with a multi-step pipeline running on Claude: retrieved sources were first summarized and clustered, and the report was then written section by section. AstaBrief writes the whole report in a single pass and skips the costly summarization and clustering steps. According to Ai2, this cuts report generation time sharply while keeping quality competitive with the earlier pipeline.

Ai2 explains that it built the model because scientists ask different things of language models than users of general-purpose chatbots do. Retrieval happens in a separate system; AstaBrief writes the report from the retrieved passages and links each claim to its source.

How it was trained

The model is built on Alibaba's open model Qwen3-8B. Ai2 applied supervised fine-tuning (SFT) on examples drawn from real users' scientific questions, filtered for quality and privacy, followed by direct preference optimization (DPO) on preference pairs. The reports in the training data were generated by several language models, including Claude and OpenAI models. Ai2 says it used statistical filters to weed out weak training examples and treated citation density, how often the text is grounded in a source, as a key signal. According to the model card, training used sequences of up to 16,000 tokens, and the model works in English only.

Alongside the model, Ai2 shared the training datasets and an intermediate checkpoint in the AstaBrief collection on Hugging Face. A sample workflow on GitHub shows how to generate reports locally from PDF files. The model runs with vLLM and Transformers, and community-made quantized versions are also on Hugging Face.

Ai2 notes that the proprietary models used to generate the training data reflected the frontier at the time and that the evaluation has not been rerun against today's strongest models. According to the institute, even when the model cites its sources it can sometimes state a claim more broadly than the source does, and human evaluation covered only a small set of questions. The model card says AstaBrief may behave inconsistently with prompt formats other than the recommended one and is intended for research and educational use. It is wise to check the sources in the generated reports before relying on them. For a fuller picture, follow evaluations from independent researchers as well.

Sources: Ai2: AstaBrief blog post, Hugging Face: AstaBrief_8B model card, Ai2: Asta, GitHub: local report workflow.