🔍 Read the full analysis: AstaBrief: Asta’s Approach To AI-Powered Report Generation on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Ai2 has released AstaBrief 8B, an open-weights model for generating reports from research questions and retrieved literature excerpts. Ai2 reports an average generation time of 51.1 seconds in Asta’s Fast mode, but has not rerun its full evaluation against current frontier models, and the supplied material does not establish report quality independently.
Ai2 has open-sourced AstaBrief 8B, a model designed to turn a research question and retrieved literature excerpts into a cited scientific report, as detailed in the original analysis. The model is also available in the institute’s Asta platform as Fast mode; Ai2 reports an average of 51.1 seconds per report, compared with 178.5 seconds for Asta’s Claude-powered Thinking mode, while saying it has not rerun its full evaluation against today’s frontier models.
AstaBrief is built on Qwen3-8B and adapted for long-form scientific synthesis, Ai2 says. The release includes the model weights, training data and an example workflow that researchers can adapt to generate reports from their own PDFs. Open weights provide the option to run or modify the model outside Asta, though the announcement does not specify whether local runs match the hosted service’s performance.
In Fast mode, the generation pipeline takes a user’s question and relevant retrieved snippets and produces a report in one pass, according to Ai2. The company says that approach skips snippet summarization and clustering, as well as section-by-section drafting used in Thinking mode. Ai2 reports full-pipeline averages of 51.1 seconds for Fast and 178.5 seconds for Thinking. Those figures make Fast about 3.5 times faster based on the reported averages, but do not by themselves show that its reports are equally accurate or complete.
Ai2 says it trained the model with supervised fine-tuning and direct preference optimization, rather than reinforcement learning. The institute describes its approach as creating and filtering examples that show desired report-writing behavior, including attention to citation grounding. The announcement does not provide enough detail to independently establish the quality comparison between the two modes.
Open Models for Research Reports
The release gives research groups access to a model they can inspect, adapt and potentially run locally, rather than relying only on a hosted report-generation service. That option may be relevant to institutions working with unpublished findings, sensitive questions or documents that they prefer not to send to an external service. Ai2’s PDF workflow also provides a starting point for using the model with researchers’ own materials.
The reported speed difference could make generated reports more convenient to create and revise as part of research work. But speed is not a measure of scientific reliability: a report can be quick to produce while omitting relevant work, misrepresenting a study’s limits or citing sources that do not support its claims. Users will need to check the evidence and citations, not just the time taken to generate the report.
Products Worth Considering
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
Amazon Kindle Scribe Colorsoft 64GB (1st Generation)
As an affiliate, we earn on qualifying purchases.
Asta’s Fast and Thinking Modes
Asta is Ai2’s platform for scientific work. The institute says its users ask questions that require comparisons across research literature, sometimes with constraints such as a particular method, population or setting. Asta’s Generate a report feature now offers AstaBrief Fast mode alongside Thinking mode, which Ai2 describes as Claude-powered.
Ai2 says its training and evaluation work relied on proprietary models that reflected the frontier at the time. Most of the work was completed in 2025, according to the institute. That timing matters when interpreting the announcement: the reported comparison is not a newly conducted benchmark against the latest available systems.
““We wanted to help scientists generate cited reports faster, with a model they could download and run themselves.””
— Ai2
Report Quality Still Needs Testing
Ai2 has not reported a full evaluation against current frontier models, and the supplied material does not explain in enough detail how report quality was measured. It also leaves open how often citations accurately support the claims beside them, how performance varies across scientific fields and question types, and whether the reported results hold when institutions run the model locally.
The reported figures concern average generation time across Asta’s pipeline; the material does not describe the hardware or configuration behind those averages. Nor does the announcement establish that Fast mode maintains the same quality as Thinking mode. These questions will require more detailed evaluations and independent testing.
Independent Testing and Further Findings
Researchers can examine the released weights and training data and adapt Ai2’s example workflow to generate reports from their own PDFs. Testing across disciplines, query types and local deployment setups could help show where the model is useful and where its summaries or citations need closer review.
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. The institute expects to share more findings from that research. A current comparison with frontier models, however, has not yet been reported.
Key Questions
What is AstaBrief 8B?
AstaBrief 8B is an open-weights model from Ai2 designed to generate cited scientific reports from a research question and retrieved literature excerpts. Ai2 says it is built on Qwen3-8B.
How fast is AstaBrief compared with Asta’s Thinking mode?
Ai2 reports average full-pipeline times of 51.1 seconds per report in Fast mode and 178.5 seconds in Claude-powered Thinking mode. These are reported averages, not a measure of report accuracy.
Can researchers run the model locally?
Ai2 has released the model weights and an example workflow, giving researchers the option to adapt the model for local use. The announcement does not establish whether local deployments perform identically to Asta’s Fast mode or specify the hardware needed.
Has AstaBrief been compared with today’s leading models?
Not in a reported full evaluation. Ai2 says most of its evaluation work was completed in 2025 and that it has not rerun the full comparison against current frontier models.
Primary source: Hugging Face · via ThorstenMeyerAI.com
