AI2 open-sources AstaBrief, a fast report-generation model for science

On October 2, the Allen Institute for AI (AI2) announced on the Hugging Face blog that it is open-sourcing AstaBrief 8B, a fast report-generation model inside Asta, its agentic platform for scientific work. The model takes a research question and retrieved literature excerpts and produces a cited report, and it is now live as Fast mode in Asta's Generate a report feature, alongside a Claude-powered Thinking mode.

Scientists use Asta with substantial context and constraints, asking it for example to compare approaches across a body of literature while accounting for a particular method, population, or setting. Many also return to generated reports later, treating them as working research artifacts rather than one-off answers.

That usage shaped the goal: help scientists generate cited reports faster, with a model they could download and run themselves. The team tested whether a small, open model trained specifically for scientific report generation could match the report quality of the proprietary models it had been using, while reducing generation time and serving costs.

According to the post, AstaBrief starts from Qwen3-8B, with most of the effort going into post-training data, evaluation, and the surrounding report-generation scaffolding. The team used tens of thousands of real research queries, applied citation-focused filtering, built preference data for DPO training, and redesigned the pipeline so the full report is written in one pass rather than section by section.

The efficiency gains are the headline. Across the full Asta pipeline, Fast mode averages 51.1 seconds per report, compared with 178.5 seconds for Thinking mode, about 3.5× faster. AI2 says the changes amount to nearly an order-of-magnitude reduction in report generation time relative to the proprietary models it tracked.

Alongside the weights, AI2 is releasing the training data and an example workflow that researchers can adapt to create reports from their own PDFs, giving them a starting point for local report generation. Open weights also let institutions run AstaBrief on their own infrastructure, which matters when research questions involve sensitive or unpublished work.

The post includes a caveat: most of the training and evaluation described was completed in 2025, so the proprietary models used to generate training data and as comparison points reflect the frontier at the time. The team has not rerun the full evaluation against today's frontier models, and it says the results are best read as evidence about particular training and system design choices.

More broadly, AstaBrief is a test case for building open language models that can be adapted to the specific demands of scientific work. The next thing to watch is whether research institutions can reproduce the results locally and keep training on top of them for their own domains.