Ai2 open-sources AstaBrief 8B, a cited-report model you self-host

The Apache-2.0 8B model turns a question plus retrieved papers into a cited report in ~51s—and Ai2 open-sourced the full retrieval-to-report stack too.

Nowline OCT 4 4:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A cited-report model that runs on your own box

    AstaBrief_8B (Apache 2.0, built on Qwen3-8B) takes a research question plus retrieved literature excerpts and writes a report with inline citations. Inside Ai2's Asta it produces a report in ~51s versus ~178s for a Claude-powered thinking mode—about 3.5x faster—while staying competitive on the SQABench-CS2 and DeepScholarBench evals. Load it with Hugging Face transformers and run it locally, no API bill.

  • The whole retrieval-to-report pipeline is open too

    Ai2 also released the training data and ai2-scholarqa-lib (Apache 2.0), the end-to-end stack that retrieves over 11M+ full texts and 100M+ abstracts, reranks, pulls quotes, outlines, and writes the cited report. It ships as a Dockerized web app plus a pip-installable package, so you can stand up your own literature-review or research-brief tool this weekend—or fine-tune AstaBrief on the open data for your own domain.