🔍 Read the full analysis: Why Asta Is Open-Sourcing Its Fast Report-Generation Model on ThorstenMeyerAI.com
Get the latest gadgets 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
TL;DR
Ai2 has open-sourced AstaBrief 8B, a model that turns a research question and retrieved literature excerpts into a cited report. Ai2 reports an average generation time of 51.1 seconds for Asta’s Fast mode, but has not rerun its full quality evaluation against current frontier models.
Ai2 has open-sourced AstaBrief 8B, a model designed to turn a research question and retrieved literature excerpts into a cited scientific report. The release includes training data and an example workflow researchers can adapt, while the model is also available in Asta’s report-generation feature as Fast mode. Ai2 reports that Fast mode averaged 51.1 seconds per report, but has not rerun its full evaluation against current frontier models.
Ai2 says AstaBrief is based on Qwen3-8B and adapted for long-form scientific synthesis. Its generation pipeline takes a user query and relevant retrieved snippets and produces a report in one pass. According to the company, training combined supervised fine-tuning with direct preference optimization; the team focused on selecting and filtering examples that demonstrated the desired report-writing behavior, including attention to citation grounding.
In Asta, Fast mode is offered alongside Thinking mode, which Ai2 describes as powered by Claude. Ai2 says Fast mode skips the snippet-summarization and clustering stages used by Thinking mode, as well as its section-by-section report writing. Across Asta’s full pipeline, the company reports an average of 51.1 seconds for Fast mode, compared with 178.5 seconds for Thinking mode. Those figures amount to roughly 3.5 times faster for Fast mode, based on the reported averages.
The release includes more than model weights: researchers can also access training data and an example workflow for generating reports from their own PDFs. Ai2 says institutions can download and run the model on their own infrastructure. That could allow some teams to keep sensitive or unpublished research materials local, though the announcement does not establish how local deployments perform compared with Asta’s hosted feature.
Speed, Access and Research Control
The release gives researchers a model and workflow they can inspect, download and adapt, rather than limiting use to a hosted service. Local deployment may appeal to universities and research teams handling unpublished work or questions they prefer not to send to an external service. The provided workflow also offers a starting point for building reports from a group’s own papers.
Ai2’s timing figures suggest the Fast mode can produce reports more quickly than the company’s Thinking mode, which could make generated summaries easier to use as working documents that researchers revisit and refine. But speed is not a measure of scientific reliability. A report still needs to represent the underlying studies accurately, preserve qualifications and limitations, and provide citations that support the claims they accompany. The supplied information does not independently establish that Fast mode matches Thinking mode in those respects.
As an affiliate, we earn on qualifying purchases.
How Asta Generates Reports
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 particular method, population or setting. AstaBrief is intended to turn a question and retrieved evidence into a longer report, rather than answer only with a short response.
The two Asta modes use different workflows, according to Ai2. Fast mode generates a report directly from retrieved snippets, while Thinking mode uses additional processing and writes the report section by section. Ai2 says it trained AstaBrief with real research queries, citation-focused filtering and preference data. It considered reinforcement learning but used supervised fine-tuning and direct preference optimization instead.
Ai2 says much of the training and evaluation work was completed in 2025, against proprietary models that reflected the frontier at that time. The company has not rerun its full evaluation against current frontier models. The reported timing comparison is specific to the Asta pipeline and does not, by itself, establish a current comparison of report quality across systems.
““We wanted to help scientists generate cited reports faster, with a model they could download and run themselves.””
— Ai2
What the Evaluation Does Not Show
The announcement does not provide a full evaluation against current frontier models; Ai2 says that comparison has not been rerun since most of its work was completed in 2025. The supplied material also does not give enough detail to independently establish the claimed report-quality comparison between Fast and Thinking modes.
Several practical questions remain open: how Ai2 measured report quality, how often citations directly support the claims they accompany, and how results vary across scientific disciplines and query types. The announcement also does not specify the hardware and configuration behind the timing averages or establish that a locally run model will perform identically to the Asta service. These are limits on what can be concluded from the release, not evidence that the model’s reports are inaccurate.
Independent Testing of AstaBrief
Researchers can examine the released weights and training data and adapt Ai2’s example workflow to generate reports from their own PDFs. Testing across fields, evidence standards and local deployment setups could clarify where the model performs well and where additional safeguards or review are needed.
Ai2 says the work forms 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 more findings. A current comparison with frontier models has not yet been reported. Until such evaluations are available, the 51.1-second figure should be read as Ai2’s reported average for Asta Fast mode, not as evidence that it produces equally reliable reports across settings.
Key Questions
What is AstaBrief 8B?
AstaBrief 8B is an open-weights model from Ai2 designed to generate a cited scientific report from a research question and retrieved literature excerpts. Ai2 says it is based on Qwen3-8B.
How fast is Asta’s Fast mode?
Ai2 reports an average of 51.1 seconds per report for Fast mode across Asta’s full pipeline, compared with 178.5 seconds for Thinking mode. The figures are company-reported averages; the announcement does not specify the hardware and configuration behind them.
Can researchers run the model locally?
Ai2 says the open weights can be downloaded and run on researchers’ own infrastructure, and the release includes an example workflow for reports based on their own PDFs. The announcement does not establish whether local performance matches the Asta service.
Has AstaBrief been compared with current frontier models?
Not in a rerun of Ai2’s full evaluation. Ai2 says most of the training and evaluation work was completed in 2025 and that it has not repeated the full comparison against current frontier models.
Does faster report generation establish report accuracy?
No. The reported timing measures how long report generation takes; it does not, on its own, establish that claims are complete or accurate or that citations support them. Those questions require evaluation of the model’s outputs.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
