Asta Opens Its Fast Report-Generation Model, AstaBrief, To The Public

🔍 Read the full analysis: Asta Opens Its Fast Report-Generation Model, AstaBrief, To The Public on ThorstenMeyerAI.com

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

Get monitors, keyboards and dev gear 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 built to generate cited reports from a research question and retrieved literature excerpts, and made it available as Fast mode in its Asta platform. The release includes model weights, training data and an example workflow researchers can adapt to produce reports from their own PDFs; Ai2 says its Fast mode averaged 51.1 seconds per report, compared with 178.5 seconds for Asta’s Claude-powered Thinking mode.

Ai2 says AstaBrief is based on Qwen3-8B and was adapted for long-form scientific synthesis. Its pipeline takes a query and relevant retrieved snippets and generates a report in one pass. According to Ai2, this bypasses steps used by Thinking mode, including snippet summarization and clustering, as well as writing the report section by section. The company attributes the faster reported generation time to this approach.

The release describes training with supervised fine-tuning and direct preference optimization. Ai2 says it focused on creating and filtering examples that demonstrate the intended report-writing behavior, including attention to citation grounding, and used real research queries and preference data. The team considered reinforcement learning but did not use it for this model, according to the source material.

Ai2 presents the 51.1-second and 178.5-second figures as averages for the full Asta report-generation pipeline. The first is about 3.5 times faster based on those reported values. However, the announcement does not supply enough information here to independently verify the figures or establish that Fast mode maintains the same report quality as Thinking mode. Generation time is not a measure of citation accuracy or scientific reliability.

A Faster Option for Research Reports

The release gives researchers access to a model they can inspect and adapt for literature-based report generation, rather than limiting them to a hosted feature. The included training data and example workflow may help research groups test the approach with their own documents and adapt it to particular questions. Ai2 says open weights also create the option of running the model on an institution’s own infrastructure, which may be relevant when queries or documents concern unpublished or sensitive research.

The reported speed difference could make it easier to generate and revisit preliminary reports as part of research work. But a shorter wait does not show that a report is accurate, complete or faithful to the underlying studies. Researchers still need to check whether citations support the statements attached to them, whether important evidence is missing, and whether the report preserves the limits of the studies it summarizes. The release offers an opportunity for testing; it does not, on the information provided, establish that those standards are consistently met.

How Asta’s Two Modes Differ

Asta is Ai2’s platform for scientific work. Ai2 says users ask it to compare approaches across research literature while applying constraints such as a specific method, population or setting. Its report-generation feature now offers AstaBrief Fast mode alongside Thinking mode, which Ai2 describes as Claude-powered.

The modes use different generation pipelines, according to Ai2. Fast mode sends the question and retrieved literature excerpts into a single report-generation step. Thinking mode uses additional stages, including organizing and summarizing snippets before producing a report section by section. Ai2 describes Fast mode as a way to reduce generation time while retaining report quality, but the supplied material does not provide sufficient detail to independently confirm that quality comparison.

Ai2 says most of the training and evaluation work was completed in 2025, using proprietary models that represented the frontier at that time. The company says it has not rerun its full evaluation against current frontier models. Its timing comparison should therefore be read as a report of the figures Ai2 supplied, not as a current head-to-head assessment of model quality.

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

— Ai2

Quality Tests Still Needed

Several performance questions remain open. The source material does not specify in detail how report quality was measured, how often citations accurately support the claims they accompany, or how performance varies by discipline and query type. Ai2 also says a full comparison with current frontier models has not been completed, so the reported results do not establish how AstaBrief compares with those systems today.

The announcement does not detail the hardware or configuration behind the reported timing averages, or establish whether a locally run version performs identically to the hosted Fast mode. Nor does it report independent evaluations of reports generated from researchers’ own PDFs. These limits matter because scientific summaries can be fast yet omit relevant evidence, misstate a study or cite material that does not support a claim.

Researchers Can Test the Release

Researchers can examine the released model weights and training data and adapt Ai2’s example workflow to generate reports from their own PDFs. Testing across research fields and deployment settings could help determine how well the model handles different evidence standards, citation requirements and local infrastructure.

Ai2 says the work is part of a broader effort to adapt open models for scientific needs, including work with scientific communities through the NSF OMAI initiative, and that it expects to share further findings. The company has not reported a new full evaluation against current frontier models. Until those results or independent tests are available, the speed figures remain Ai2’s reported averages, while the quality and citation reliability of outputs remain matters for further assessment.

Key Questions

What is AstaBrief 8B?

AstaBrief 8B is Ai2’s open-weight model for generating reports from a research question and retrieved literature excerpts. Ai2 says it is based on Qwen3-8B and adapted for long-form scientific synthesis.

How fast is AstaBrief compared with Thinking mode?

Ai2 reports an average of 51.1 seconds per report for Fast mode and 178.5 seconds for Asta’s Claude-powered Thinking mode. These are company-reported averages for the full pipelines; the source material does not give enough detail to independently verify the comparison.

Can researchers run AstaBrief locally?

Ai2 has released open weights and an example workflow that researchers can adapt, including for reports based on their own PDFs. Ai2 says open weights give institutions the option of local deployment, but the supplied material does not establish that local performance matches the hosted service.

Has Ai2 shown that AstaBrief reports are as reliable as Thinking mode?

The supplied information does not establish that. Ai2 describes Fast mode as aiming to maintain report quality while reducing generation time, but details of the quality measures and citation accuracy are not provided here. Ai2 also says it has not rerun its full evaluation against current frontier models.

Primary source: Hugging Face · via ThorstenMeyerAI.com