Generate research paper reports locally: Ai2 releases AstaBrief weights

One pass. The existing Claude-based Thinking mode first summarizes and groups excerpts, then writes the report section by section. AstaBrief in Fast mode generates the report in one pass. According to Ai2's measurements on Oct 2, the entire process takes 3.5 times less time on average.

| Asta mode | Average report generation time | | --- | --- | | Fast mode | 51.1 seconds | | Thinking mode on Claude | 178.5 seconds |

No installation needed. In Asta, open Generate a report and select Fast mode. For reports based on your own PDFs, Ai2 suggests adapting the example from ai2-scholarqa-lib: a ready-made document processing example for your own application.

To run it yourself, the weights are available in the allenai/AstaBrief_8B repository on Hugging Face. In the sft_prompt.txt template, fill [QUERY] with a research question and [SECTION_REFERENCES] with paper excerpts and citation identifiers. The model receives prepared material rather than a folder of PDFs directly.

Ai2 trained AstaBrief on Qwen3-8B. In the ScholarQA-CS2 test with 100 computer science questions, the average score rose from 77.3 for the base model to 83.7 after training on examples and 87.0 after preference training. Citation accuracy rose from 76.2 to 90.5.

Of the 374 users who tried Fast mode, 29.1% used it on at least two days. Another metric from Ai2: 23% continued using this mode and stopped choosing Thinking mode for new discussions.

Original source: [Ai2 released AstaBrief weights and Fast mode measurements, Oct 2, 2026](https://huggingface.co/blog/allenai/astabrief).

Resources for getting started: [Asta with Fast mode selection](https://asta.allen.ai/), [AstaBrief 8B weights](https://huggingface.co/allenai/AstaBrief_8B), [example for your own PDFs](https://github.com/allenai/ai2-scholarqa-lib/tree/main/api/scholarqa/lite).

Ai2 released the weights under Apache 2.0, so AstaBrief can be integrated into your own report generation service.