Open-sourcing AstaBrief, The Fast Report-generation Model In Asta

TL;DR

Prime Big Deal Days · Oct 6–7Offer from Amazon

Get tech for your team delivered free — and shop member deals

• Fast, free delivery on millions of items
• Access to Prime Big Deal Days deals on October 6–7
• Prime Video, Amazon Music and more included

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Ai2 has open-sourced AstaBrief 8B, a model designed to turn research questions and retrieved literature excerpts into cited reports, along with its training data and an example local workflow. Ai2 reports that Asta’s Fast mode averages 51.1 seconds per report, versus 178.5 seconds for its Claude-powered Thinking mode; the comparison reflects evaluations largely completed in 2025, not current frontier models.

Ai2 has open-sourced AstaBrief 8B, a model built to produce cited scientific reports from a research question and retrieved literature excerpts. The model is available in Asta’s Fast mode, and Ai2 says it is releasing the weights, training data and an example workflow so researchers can study the approach and run report generation on their own infrastructure.

Ai2 reports that Asta Fast mode takes an average of 51.1 seconds per report across the full Asta pipeline. Its Claude-powered Thinking mode averages 178.5 seconds, making Fast mode about 3.5 times faster in the company’s comparison. Those figures describe Ai2’s tested system and are not an independent benchmark of the model alone.

AstaBrief starts from Qwen3-8B. Ai2 says it focused on supervised fine-tuning and direct preference optimization, alongside the selection and filtering of training examples, rather than using reinforcement learning for this model. The report pipeline was also redesigned to generate the full report in one pass from the query and relevant excerpts, bypassing stages used in Thinking mode, including snippet summarization and clustering.

The release includes model weights and training data, as well as an example workflow for generating reports from researchers’ own PDFs. Ai2 presents local deployment as useful for institutions handling sensitive or unpublished research questions. The source post describes the workflow as a starting point for adaptation; it does not establish that every institution can deploy it without additional technical work.

Faster Reports, Local Research Workflows

Making the model and data available gives research groups a way to examine how an open model can synthesize literature, rather than relying only on a hosted proprietary service. If the workflow can be adapted to a group’s own documents and infrastructure, it may also help keep sensitive research queries and materials within that institution’s environment. Ai2 says the release is intended to let others study, reproduce and build on its report-generation approach.

The reported speed difference matters for a workflow in which users ask questions with substantial constraints, then revise or revisit reports. A shorter wait could make generating an initial report more practical. But the comparison is about generation time in Asta’s full pipeline; the source does not provide a cost figure or establish that Fast mode is more accurate than Thinking mode. Nor does speed by itself show that a report faithfully represents every study’s findings.

Scientific reports need traceable support as well as fluent writing. Ai2 says it designed the training and filtering process around citation grounding and preserving what the evidence supports. The release lets other researchers inspect parts of that work, though the source material does not establish that the model eliminates citation errors or unsupported conclusions.

How AstaBrief Fits Into Asta

Asta is Ai2’s platform for scientific work. Its Generate a report feature uses a research question and retrieved literature to produce a cited synthesis. AstaBrief is offered as Fast mode alongside a Claude-powered Thinking mode. According to Ai2, the two modes use different report-generation pipelines: Fast produces the complete report in one pass, while Thinking mode includes additional processing stages.

Ai2 describes the work as an investigation into whether a small open-weights model, trained specifically for scientific report generation, could approach the report quality of proprietary models while reducing time and serving costs. The team says it drew on real research queries and used citation-focused filtering and preference data. The supplied source excerpt does not give the full details of the query collection, data quantities, evaluation procedures or release terms, so those details should not be inferred here.

The timing of the evaluation limits what the results establish. Ai2 says most training and evaluation were completed in 2025, and that the proprietary models used to produce training data and as comparison points reflected the frontier at that time. The company says it has not rerun the full evaluation against current frontier models. Its performance and timing claims therefore describe the systems and tests it used, not a current, like-for-like comparison with every available model.

“We wanted to help scientists generate cited reports faster, with a model they could download and run themselves.”

— Ai2, in its Hugging Face announcement

Limits of the Published Comparison

The announcement does not report a new evaluation against today’s leading proprietary models. Ai2 explicitly says the full evaluation has not been rerun against current frontier models, and that most work was completed in 2025. The reported timing comparison also does not, by itself, answer whether the modes produce equivalent results on every research question or field.

Further details are also needed to judge how well the released system handles weak or conflicting evidence, citation accuracy, and studies whose findings do not support a broad conclusion. Ai2 describes its training as focused on grounding and evidence preservation, but the supplied material does not establish error rates, independent validation, or performance across scientific disciplines. The release’s exact license and the full scope of its training-data documentation are not specified in the source excerpt.

Researchers Can Test Local Runs

Ai2 has made the model, training data and an example workflow available for researchers to examine and adapt. The practical next step is for research groups to test the workflow with their own papers and questions, checking report quality, citations, runtime and any infrastructure requirements. Ai2 says the example workflow is intended as a starting point for generating reports from local PDFs.

Further published evaluations would help clarify how AstaBrief performs across fields and how it compares with newer systems. Ai2 says it plans to share more findings from its broader work with scientific communities, but the announcement gives no date for a follow-up evaluation or a specific next release.

Key Questions

What is AstaBrief?

AstaBrief 8B is an open-weights model from Ai2 designed to generate cited scientific reports from a research question and retrieved literature excerpts. It is available as Fast mode in Asta’s report-generation feature.

How much faster is Asta Fast mode?

Ai2 reports an average of 51.1 seconds per report for Fast mode and 178.5 seconds for its Claude-powered Thinking mode across the full Asta pipeline. That is about 3.5 times faster in the reported comparison; it is not a standalone model benchmark.

What has Ai2 released?

Ai2 says it is releasing AstaBrief’s model weights and training data, along with an example workflow that researchers can adapt to create reports from their own PDFs. The source announcement does not specify all release terms in the supplied material.

Are the results a comparison with current AI models?

No. Ai2 says most training and evaluation were completed in 2025 and that it has not rerun the full evaluation against today’s frontier models. The reported findings apply to the systems and tests used at that time.

Source: rss